Professional Documents
Culture Documents
In 2004 the AIChE Journal celebrated its 50th anniversary. In connection with this milestone, a symposium on The Future of Chemical
Engineering Research was organized at the Annual AIChE Meeting in Austin. The symposium consisted of three sessions entitled:
Fundamentals
Biological Engineering
Complex Systems
In each session seven talks by leading chemical engineers were followed by a three-member panel discussion. A one-person
overview that integrated the talks and discussions within the context of our disciplines research tradition closed the proceedings.
The symposium proved an excellent opportunity for taking a broad look at the current landscape of chemical engineering
research, and for thinking about its future. It is clearly impossible to organize three sessions that can provide a comprehensive
representation of every important aspect of chemical engineering research. Nevertheless, we believe that the scope and quality of
the talks were such that it is of value to our profession to preserve a written record of the symposium. We are delighted that the
AIChE Journal will be the medium for its publication.
This issue of the AIChE Journal features the third collection of Perspective articles dedicated to the 2004 Symposium The
Future of Chemical Engineering Research. The three articles in this issue correspond to the session on Biological Engineering.
The session on Complex Systems appeared in July, and the series of Fundamentals Perspectives appeared in the September issue.
We are grateful to the many speakers who were able to contribute to this collection by putting their talks and thoughts on paper.
We feel honored and privileged to have organized the symposium and to edit this collection of papers. The program of the session
on Biological Engineering is reproduced below.
3082
December 2005
AIChE Journal
Perspectives
AIChE Journal
December 2005
3083
Perspective
op-down exercises in research area prognostication often miss the mark. The most general predictions can be
platitudinously unfalsiable; the most specic are often
amusingly wrong in retrospect. Given the primacy of investigator-initiated efforts in the successful denition and exploitation of new research areas, it is fair to question the utility of
the future of. . . pieces, such as this one.
I well recall sitting through keynote presentations by eminent
senior colleagues in the biochemical engineering eld at national
meetings in the late 1980s and early 1990s, the tenor of which was
that the essential premise of my research program was an enormous mistake, straying too far from the hard ChE core. A particularly memorable assertion was that It is better to be a rst-rate
chemical engineer than a second-rate biologist. (Fortunately
these do not appear to be the only two available options.) Editorializing is an ineffective means of squelching initiative, by comparison to the ruthless efciency of the existing marketplace of
ideas. Successful research directions are determined, over a sufcient time span, by peer-reviewed funding and publication, the
interest of new students, and the willingness of academia and
industry to hire students trained to perform such research. If
exercises such as this one are capable of serving a useful purpose,
perhaps it is to constructively recognize incipient grass-roots
movements and highlight exciting challenges.
3084
December 2005
role in biological function and many pathologies. The therapeutic efcacy of a protein drug is inextricably linked to its
binding properties, and consequently tools for engineering protein binding are indispensable, as well as analytical tools to
properly determine biophysical design criteria.1
At present the most powerful and robust approach to engineering protein properties is Directed Evolution, and the eld
of chemical engineering has quietly gone about becoming the
predominant academic home for Directed Evolution research.
A cursory examination of departmental web sites reveals over
20 ChE faculty in the U.S. (most of whom began their careers
in the past 15 years) whose research programs are extensively
dedicated to developing or using Directed Evolution. Combinatorial polypeptide library screening consists of a series of
equilibrium and/or transient biochemical transformations that
are well suited to analysis with classical chemical reaction
engineering tools.2-5 Widely-practiced innovations in protein
screening methodology have originated from ChE research
groups.6-8 The fundamental structure of the tness landscape,
and improved search methods for it, has been a fruitful source
for interesting research problems.9-14 A growing direction is the
use of chemical engineering analyses to guide the development
of optimized protein biopharmaceuticals.1, 15-19 In terms of
analytical contributions, cellular signaling pathways,20, 21 metabolism,22, 23 and immunology24, 25 hold tremendous promise
for chemical engineers, and these approaches are likely to
contribute to rationalization of protein drug pharmacology.
AIChE Journal
equilibria, and the rates of biomolecular movement by diffusion and convection. The ChE triumvirate core of kinetics,
thermo, and transport are the required tools for such analyses.
Although BME curricula generally include transport coursework, BME training programs in general do not incorporate
kinetics and chemistry to the extent necessary for modern
biomolecular engineering.
ChE should not rest on its laurels, however; incorporating
biology into the ChE curriculum is not so simple as adding a
few new examples or homework problems, because there are
new intellectual principles from biochemistry, biophysics, and
cell biology that must be integrated throughout the toolkit. The
most effective means for syllabus evolution will be to staff
courses with champions dedicated to change; given the large
numbers of young ChE faculty with biological interests now
disseminated throughout the U.S., this is already occurring.
These efforts will be amplied by the emergence of new
textbooks incorporating biological principles in an integrated
fashion rather than grafted onto existing outlines.
Literature Cited
1. Rao BM, Lauffenburger DA, Wittrup KD. Integrating celllevel kinetic modeling into the design of engineered protein therapeutics. Nat Biotechnol. 2005;23(2):191-4.
2. Maheshri N, Schaffer DV. Computational and experimental analysis of DNA shufing. Proc Natl Acad Sci U S A.
2003;100(6):3071-6.
3. Daugherty PS, Olsen MJ, Iverson BL, Georgiou G. Development of an optimized expression system for the screening of antibody libraries displayed on the Escherichia coli
surface. Protein Eng. 1999;12(7):613-21.
4. Boder ET, Wittrup KD. Optimal screening of surface-displayed polypeptide libraries. Biotechnol Prog. 1998;14(1):5562.
5. Moore GL, Maranas CD. Modeling DNA mutation and
recombination for directed evolution experiments. J Theor
Biol. 2000;205(3):483-503.
6. Zhao H, Giver L, Shao Z, Affholter JA, Arnold FH. Molecular evolution by staggered extension process (StEP) in vitro
recombination. Nat Biotechnol. 1998;16(3):258-61.
7. Francisco JA, Earhart CF, Georgiou G. Transport and anchoring of beta-lactamase to the external surface of Escherichia
coli. Proc Natl Acad Sci U S A. 1992;89(7):2713-7.
8. Boder ET, Wittrup KD. Yeast surface display for screening combinatorial polypeptide libraries. Nat Biotechnol.
1997;15(6):553-7.
9. Voigt CA, Mayo SL, Arnold FH, Wang ZG. Computational
method to reduce the search space for directed protein evolution. Proc Natl Acad Sci U S A. 2001;98(7):3778-83.
10. Moore GL, Maranas CD. Identifying residue-residue clashes
in protein hybrids by using a second-order mean-eld approach. Proc Natl Acad Sci U S A. 2003;100(9):5091-6.
AIChE Journal
December 2005
11. Bogarad LD, Deem MW. A hierarchical approach to protein molecular evolution. Proc Natl Acad Sci U S A.
1999;96(6):2591-5.
12. Earl DJ, Deem MW. Evolvability is a selectable trait. Proc
Natl Acad Sci U S A. 2004;101(32):11531-6.
13. Daugherty PS, Chen G, Iverson BL, Georgiou G. Quantitative analysis of the effect of the mutation frequency on
the afnity maturation of single chain Fv antibodies. Proc
Natl Acad Sci U S A. 2000;97(5):2029-34.
14. Drummond DA, Iverson BL, Georgiou G, Arnold FH.
Why high-error-rate random mutagenesis libraries are enriched in functional and improved proteins. J Mol Biol.
2005;350(4):806-16.
15. Graff CP, Wittrup KD. Theoretical analysis of antibody
targeting of tumor spheroids: importance of dosage for
penetration, and afnity for retention. Cancer Res. 2003;
63(6):1288-96.
16. Sarkar CA, Lauffenburger DA. Cell-level pharmacokinetic
model of granulocyte colony-stimulating factor: implications for ligand lifetime and potency in vivo. Mol Pharmacol. 2003;63(1):147-58.
17. Rao BM, Driver I, Lauffenburger DA, Wittrup KD. HighAfnity CD25-Binding IL-2 Mutants Potently Stimulate Persistent T Cell Growth. Biochemistry. 2005;44(31):10696-701.
18. Haugh JM. Mathematical model of human growth hormone (hGH)-stimulated cell proliferation explains the efcacy of hGH variants as receptor agonists or antagonists.
Biotechnol Prog. 2004;20(5): 1337-44.
19. Kim JR. Gibson TJ, Murphy RM. Targeted control of kinetics
of beta-amyloid self-association by surface tension-modifying peptides. J Biol Chem. 2003;278(42):40730-5.
20. Kao KC, Yang YL, Boscolo R, Sabatti C, Roychowdhury
V, Liao JC. Transcriptome-based determination of multiple transcription regulator activities in Escherichia coli by
using network component analysis. Proc Natl Acad Sci U
S A. 2004;101(2):641-6.
21. Sachs K, Perez O, Peer D, Lauffenburger DA, Nolan GP.
Causal protein-signaling networks derived from multiparameter single-cell data. Science. 2005;308(5721):523-9.
22. Price ND, Reed JL, Palsson BO. Genome-scale models of
microbial cells: evaluating the consequences of constraints. Nat Rev Microbiol. 2004;2(11):886-97.
23. Alper H, Miyaoku K, Stephanopoulos G. Construction of
lycopene-overproducing E. coli strains by combining systematic and combinatorial gene knockout targets. Nat Biotechnol. 2005;23(5):612-6.
24. Chakraborty AK. Decoding communications between cells
in the immune system using principles of chemical engineering. AIChE J. 2003;49(7):1614-1620.
25. Munoz ET, Deem MW. Epitope analysis for inuenza
vaccine design. Vaccine. 2005;23(9):1144-8.
26. Brekke OH, Sandlie I. Therapeutic antibodies for human
diseases at the dawn of the twenty-rst century. Nat Rev
Drug Discov. 2003;2(1):52-62.
27. Vasserot AP, Dickinson CD, Tang Y, Huse WD, Manchester
KS, Watkins JD. Optimization of protein therapeutics by
directed evolution. Drug Discov Today. 2003;8(3):118-26.
28. Lazar GA, Marshall SA, Plecs JJ, Mayo SL, Desjarlais JR.
Designing proteins for therapeutic applications. Curr Opin
Struct Biol. 2003;13(4):513-8.
Vol. 51, No. 12
3085
Perspective
Executive Summary
3086
December 2005
AIChE Journal
December 2005
3087
mune disease (Leirisalo-Repo, 2005; Kaplan et al., 1997). However, the strength and signicance of this correlation is controversial (Carty et al., 2003). Our model suggests a broad distribution
for the time of onset of autoimmune disease due to chronic
infection. Researchers have been looking for a clear, signicant
correlation in time, but a long distribution of onset times would
lead to weaker statistical correlations, particularly in those cases
where the infection persisted the longest. Searching for this distribution could elucidate this immunological puzzle and settle the
scientic controversy.
We found that the human immune system evolved to minimize the risk of cross-reactivity. For example, each cell in our
bodies contains about 100,000 proteins with an average of 500
amino acids apiece. Consequently, there are about 1012 potential docking sites, or epitopes, where antibodies could mistakenly attach themselves to proteins in a healthy cell. The mutation response method employed by our adaptive immune
system seems keyed to this number, producing antibodies that
are statistically likely to mistakenly bond with healthy proteins
slightly less than one in 1012 times, meaning that on average,
they recognize only invading pathogens.
Randomness of Cancer
The percentage of Americans dying from cancer is the
same as what it was in 1970. . . and the same as what it was
in 1950 (Leaf, 2004). Although some progress has been
made, especially for childhood cancers, cancer remains a
largely unsolved purge of modern society. Mouse models
remain largely unpredictive, and cancer seems a tremen3088
December 2005
AIChE Journal
Figure 3. Value of the parameter mixing round for vaccination to different lymph nodes at different
distances from the heart.
Humans have several hundred lymph nodes. For effective
polytopic vaccination, well-separated sites on different limbs
are used.
Summary
By a discussion of three examples, I hope to have convinced the
reader that signicant unsolved theoretical problems exist in medicine and that statistical mechanics has a pivotal role to play in
their solution. The importance of randomness to the proper functioning of the immune system seems an especially ripe topic for
statistical mechanical analysis. Many, if not most, problems in
immunological diversity remain open, and identication of the
models and theories to tackle these problems is just starting.
Interaction with immunologists and pathologists has proved helpful to my group as we explore these issues. We are also fortunate
as a eld and as a profession that graduate students are keen to
contribute to the new mathematics of biology.
Acknowledgments
It is a pleasure to acknowledge the group members who
contributed to this work: Enrique T. Munoz, David J. Earl,
Vishal Gupta, Hao Zhou, Jun Sun, Ming Yang, and Jeong-Man
Park. Support by the National Institutes of Health of research in
my group is gratefully acknowledged.
Literature Cited
Allakhverdov, A., and M. Enserink, Bird Flu Hits Russia,
Science, 309, 861 (2005).
Ault, A., Shifting Tactics in the Battle Against Inuenza,
Science, 303, 1280 (2003).
Carty, S. M., N. Snowden, and A. J. Silman, Should Infection
Still be Considered as the Most Likely Triggering Factor for
Rheumatoid Arthritis? J. Rheumatology, 30, 425 (2003).
AIChE Journal
December 2005
3089
3090
December 2005
AIChE Journal
Perspective
n a recent AIChE Journal perspective on Metabolic Engineering1 we asked rhetorically whether the microbial world
is so diverse as to allow one to isolate some microbe
capable of producing any desired molecule. It turns out that this
is very likely the case, a manifestation of the enormous diversity of molecules and reaction processes resident in a microbe.
These actually constitute the mechanisms by which cellular
functions are being carried out. The problem with most organisms is that they may make only traces of any single desired
molecule and under conditions that may be difcult to implement on an industrial scale. These microbes must be improved
before their potential is realized. Furthermore, after demonstrating that it is possible to cross the species barrier, pathways
that are incomplete for the production of a molecule in one
organism can now be completed by transferring the missing
pieces from another microbe. Thus, the enormous diversity
present in an array of species can now be commandeered for
accomplishing a specic purpose, such as the production of
entirely new products or the construction of new synthesis
routes for existing products. This is the goal and essence of
metabolic engineering.
Metabolic engineering was developed in the previous decade
to improve industrial strains using modern genetic tools. While
microorganisms were modied before by random mutation and
selection methods, the development of recombinant technologies in early 80s allowed directed strain modication by introducing specic genes conferring desirable properties to cells
for industrial, medical and environmental applications. Metabolic engineering, thus, emerged as the scientic discipline
occupied with the improvement of cellular properties through
introduction to cells of specic transport, enzymatic or regulatory reactions using, primarily, recombinant technologies.2,3
In less than a decade an impressive number of metabolic
engineering applications have appeared in diverse areas includ*Correspondence concerning this article should be addressed to G. Stephanopoulos
at gregstep@mit.edu.
AIChE Journal
December 2005
3091
December 2005
Acknowledgment
NSF award number: BES-0 331364.
Literature Cited
1. Stephanopoulos G. Metabolic engineering: Perspective of
a chemical engineer. AIChE J. 2002:48;920-926.
2. Bailey JE.Toward a science of metabolic engineering. Science, 1991:252;1668.
3. Stephanopoulos G,Vallino JJ. Network rigidity and metaVol. 51, No. 12
AIChE Journal
4.
5.
6.
7.
8.
9.
10.
11.
12.
AIChE Journal
December 2005
13.
14.
15.
16.
17.
18.
19.
20.
21.
3093
Mixing times, power consumption, heat transfer, and scale-up predictions in an agitated
vessel require the use of correlations between dimensionless groups such as Prandtl, Nusselt,
Power, and Reynolds numbers. These dimensionless numbers are now well established for an
agitated vessel equipped with a vertically and centrally mounted impeller in the tank for both
Newtonian and non-Newtonian uids. To our knowledge, there is more ambiguity as to the
denition of the characteristic speed and dimensions, which should be taken into account in
the dimensional analysis of planetary mixers. The aim of this paper is twofold: (1) to propose
modied Reynolds and Power numbers for planetary mixers and (2) to ascertain the reliability
of the modied dimensionless number proposed for a particular planetary mixer, The TRIAXE, which uses a combination of rotation and gyration of a pitched blade turbine to achieve
mixing. The modied Reynolds and Power numbers proposed involve the maximum tip speed
as characteristic velocity and are consistent with the denition of traditional Reynolds and
Power numbers when only a single revolution around the vertical axis of the mixing device
occurs in the vessel, as is the case for a standard mixing system. Experimental power
measurements carried out with a planetary mixer when mixing highly viscous Newtonian
uids show that the modied Reynolds and Power numbers proposed succeed in obtaining a
unique power curve for the mixing system independently of the speed ratio. This close
agreement proves that the modied Reynolds and Power numbers are well adapted for
engineering purposes and can be used to compare the power-consumption performances of
planetary mixers with well-established technologies. 2005 American Institute of Chemical
Engineers AIChE J, 51: 3094 3100, 2005
Introduction
In the second half of the twentieth century, the systematic
use of dimensional analysis to investigate mixing processes has
allowed this eld to evolve from arts into sciences. Today the
whole eld of classical stirring technology (here, the word
classical refers to impellers vertically and centrally mounted in
the tank) has been examined, so that the denition of signicant dimensionless groups has now become well established.
Correspondence concerning this article should be addressed to G. Delaplace at
delapla@lille.inra.fr.
3094
December 2005
AIChE Journal
Experimental
Mixing equipment
The mixing equipment used in this investigation is the TRIAXE system (HOGNON S.A., Mormant, France), which
allows the agitator to combine two motions: gyration and
rotation (Figure 1). This planetary mixer is characterized by
two revolutionary motions that are nearly perpendicular. Gyration is a revolution of the agitator around a vertical axis,
whereas rotation is a revolution of the agitator around a nearly
horizontal axis. These double motions allow the agitator to
periodically come in contact with the entire volume of the
vessel. In this work, the mixing tool for the TRIAXE system
is a four pitched blade turbine (Figure 2). The tank used is a
transparent glass cylinder with rounded bottom (Figure 2). The
diameter of the vessel is 0.4 m. In this work, experimental
measurements were carried out when the agitator was fully
immersed in the liquid. This liquid height corresponds to a
liquid volume of 38 L, corresponding to a liquid height of 0.39
m.
The mixing equipment was driven by two variable-speed
motors. To obtain the total power consumption of the TRIAXE system, power draw measurements were carried out
alternately for the two variable-speed drive motors that control
the impeller revolutions. To do so, a torque meter (Scaime Inc.,
Annemasse Cedex, France) ranging from 0 to 5 Nm, was
mounted alternately on the two motor drive shafts (Figure 2)
and the torque was measured for various impeller speed ratios.
Total power draw was estimated as a simple summation of the
two motor drives. More exactly, the procedure to obtain power
requirement of each motor was the following:
AIChE Journal
December 2005
3095
Figure 2. Picture of the agitator and vessel of the TRIAXE system used in this investigation.
Red circle on the picture refers to the torquemeter device. The
vessel tank diameter T and liquid height HL were set equal to
0.4 and 0.39 m, respectively. For the planetary mixer studied,
ds 0.14 m and D 0.38 m.
Agitated uid
The agitated uid is a highly viscous Newtonian uid. The
liquid consists of glucose syrup/water mixtures of various
viscosities ranging from 15 to 29 Pas, depending on the
temperature encountered in the vessel. The rheological properties and densities of the test uid were obtained at the same
temperatures as those encountered in the mixing equipment.
Density at 20C of the test uid is 1400 kg/m3. The rheological
properties of the test uid were measured in standard controlled
rotational speed concentric cylinders (Rheomat 30, Contraves
AG, Zurich, Switzerland). The shear rate range applied to the
controlled shear rate viscosimeter varied from 0 to 500 s1and
corresponds to the shear rate range encountered in the tank.
Theory
Dimensional analysis for an agitated vessel equipped
with an impeller vertically and centrally mounted in the
tank
3096
December 2005
When mixing Newtonian liquids in an agitated vessel without bafes, equipped with an impeller vertically and centrally
mounted in the tank (Figure 4), the power P of a given stirrer
type and given installation conditions (vessel diameter T, agitator height H, liquid height HL, and bottom clearance C) in a
homogeneous liquid depends on the agitator diameter d (as the
characteristic length), the material parameters of the liquid
Vol. 51, No. 12
AIChE Journal
F 5 N pG, ReG ,
NR D
,
0
NG ds
(5)
3 5
where N p G P/ N G
d s and ReG N G d s2 /.
For a given planetary system, the ratio D/ds is xed and the
power characteristic is reduced to a mutual dependency between each of the following three parameters constituting the
pi-set
F 6 N pG, ReG ,
(1)
(2)
NR
0
NG
(6)
(7)
(3)
December 2005
3097
u impeller tipt
2NG
2NR
D
2
2NG
ds
2
D 2 2
cos 2NR t 22 NR NG ds D sin2NR t
2
1/ 2
(10)
(8)
(9)
To sum up, the dimensional analysis of the power consumption of the TRIAXE system in a xed installation condition
when mixing highly viscous uids, leads either to a relationship between three pi-numbers (Eq. 6) or two pi-numbers (Eq.
8) if a characteristic speed proportional to the maximum impeller tip speed is introduced in the parametric dimensional
space. In addition, it has been shown that the pi-numbers can be
reduced to well-known Reynolds and Power numbers, when
the rotational speed is zero. This is quite logical because, in this
case, the TRIAXE system is transformed to a classical mixing
system equipped with an agitator vertically and centrally
mounted in the tank.
In the following section, the way to compute the characteristic speed for the TRIAXE system will be detailed. Then, the
reliability of the two pi-numbers proposed will be ascertained,
using power consumption measurements.
December 2005
AIChE Journal
Acknowledgments
The authors are grateful to F. Brisard and J. F. Dauphin for valuable
work in obtaining the experimental data and gures presented here.
Literature Cited
Conclusion
In this paper, modied Reynolds and Power numbers for a
planetary mixerthe TRIAXE system, which combines dual
revolutionary motions has been developed.
It has been shown that the proposed modied Reynolds and
Power numbers, which involve the maximum impeller tip
speed as characteristic velocity and a dimension perpendicular
to the vertical axis of revolution as characteristic length, allow
one to obtain a unique power characteristic of the mixing
system, regardless of the speed ratios. Moreover, the characteristic length and velocity chosen ensure that the modied
dimensionless numbers proposed are consistent with the denition of traditional Reynolds and Power numbers, when the
impeller performed only a single revolution around the vertical
axis in the tank, as in the case of a classical mixing system. So,
it clearly appeared that the modied Reynolds and Power
numbers can be easily used to compare the power consumption
of planetary mixers with that of conventional mixing systems.
In our judgment, the suggested modied Power and Reynolds numbers can be used for other planetary mixers that
combine two revolutions around a vertical axis. Only the
expression of the characteristic velocity must be modied, to
allow for the differing operating conditions. In contrast to
double-armed planetary mixers or twin-blade arrangement, dimensional analysis dened in Eq. 6 should be preferred because there is no way to reduce the pi-numbers using a charAIChE Journal
December 2005
Appendix A Scheme 1
Considering M to be a point located at the TRIAXE impeller tip, the absolute velocity of M in the rotating reference
frame (e x , e y , e z ) is given by
V a
dOM
dt
dOO
dt
dOO
dt
dOO
dt
OO
dOM
dt
with
R/R
Scheme 1.
Vol. 51, No. 12
3099
dOO
dt
OM
OO
and
dOM
dt
dOM
dt
R
OM
R/R
with
dOM
dt
V a
D/2cos2NR t
OM
D/2sin2NR t
0
0
D/2cos2NR t
0
0
2NG D/2sin2NR t
0
0
2NG D/2cos2NR t
0
0 2NG ds /2
2
0
G
0
ds /2
0
D/22NR sin2NR t
D/22NR cos2NR t
0
Appendix B
It can be shown that the function t x uimpeller tip(t), dened in
Eq. 10, reaches an extremum when the derivative of the function
t x w(t) is equal to zero with w(t) given by w(t) 2(D/
2)2NR2NG(ds/2)sin(2NRt) [2NG(D/2)]2cos2(2NRt).
The derivative of the function t x w(t) is dw(t)/dt
2NRcos(2NRt){[2(D/2)2NR2NG(ds/2)] 2[2NG(D/
2)]2sin(2NRt)}.
Consequently dw(t)/dt 0 when the following conditions
apply:
t
1
4N R
3
1
,
4N R 4NR
3
1
,
4N R 4NR
or at
or
arcsin
3
t
4N R
NR ds
,
NG d
arcsin
NR ds
NG d
NR ds
sin2NR t
NG d
3100
December 2005
AIChE Journal
Introduction
Particle transport by uids is a common process in nature
and in technical applications. Examples of such two-phase
ows are mixing in bioreactors, the spreading of small particles
in the atmosphere, transport of particles through pipes and
channels, separation in cyclones and sediment transport in
rivers and oceans. In some cases, such as pneumatic conveying,
the transport of particles is strongly inuenced by turbulence.
Thus, turbulent dispersion must be accounted for when the
distribution of particles is calculated. Studies of pneumatic
conveying show that there is a minimum mean velocity at
which particles can be transported through a horizontal pipe or
a channel without being deposited on the pipe bottom. Predictions for this minimum velocity have been derived and discussed by, for example, 1 4. At lower velocities the turbulent
velocity will not be high enough to carry the particles and they
will accumulate at the lower pipe wall. This indicates that, to
avoid this accumulation, turbulent dispersion must be great
enough to counteract the inuence of gravity on the particles.
When numerical calculations are made of these kinds of ows,
Correspondence concerning this article should be addressed to R. Rundqvist at
robert.rundqvist@fcc.chalmers.se
AIChE Journal
December 2005
3101
Experimental
Experimental setup
Experimental setup is shown in Figure 1. The particle volume fraction prole was measured in a horizontal aluminum
pipe 10 m long and 0.108 m in dia. A classical Venturi tube
with a machined convergent section with a dia. ratio of 0.75,
designed according to the ISO 5167-1 standard, was used to
measure the mean velocity. At low particle concentrations the
extra pressure drop over the Venturi tube due to particles in the
ow is negligible. The suction fan was set to give a mean gas
velocity of 12 m/s. As particles were introduced, the gas
velocity dropped and the fan was adjusted to keep the original
velocity. The particles were injected at the inlet with a loading
ratio of 0.1. The particles used were poly-acrylate particles
having a diameter of 0.4 mm. The density of the particles is
1,000 kg/m3. An optical probe, described in the next section,
was traversed vertically along the pipe diameter, close to the
end of the pipe. Measurements were made both from the top
and the bottom of the test section in order to get a prole for the
entire pipe diameter. The probe introduces incoherent infrared
light into the measuring volume by an optical ber. Another
ber adjacent to the introducing ber, slanted by 45 degrees to
limit the measuring volume, is used to receive the reected
light from the particles. The conguration of the bers in the
probe is shown in Figure 2.
Optical probe
December 2005
AIChE Journal
Numerical
Two-uid model
The Eulerian two-uid model used in this work is based on
the traditional forms of the continuity and momentum equations, cf. for example, 29 31. The basic form of the Reynolds
decomposed and volume averaged continuity and momentum
equations are
U 0
t k k
x j k k kj
K(U pi U gi) if k g
d
K(U U ) if k p
M kIi
gi
pi
(3)
For a dispersed ow, the generalised drag force per unit volume in a suspension of particles is written
d
n pF idr
M kIi
(4)
p
Vp
(5)
F sp
d p2
1 sp
C d gU rU r
2
4
(6)
24
(1 0.15(Red )0.687 ) if Red 1000
C Red
0.44
if Red 1000
sp
d
(7)
Here, Red is the Reynolds number based on the particle diameter and the relative velocity between the phases.
The viscous force in the momentum equation is given by
(1)
V ki
d
U U k
P MkIi
U
t k k ki
x j k k ki kj
xj ij
k k gi Cki Vki
(2)
where the phase index k is g for the gas phase and p for the
particle phase, k is the average volumetric occurrence of
phase k, k is the density, Uki is the velocity, P is the
d
gas-phase pressure, MkIi
is the generalised drag, gi is the
gravitational acceleration, Vki is the viscous force, and Cki
represents the particle-particle interaction force, which is
zero in the gas phase. These are the laminar forms of the
transport equations. The equations including the additional
terms that take turbulence into account are presented below
in this paragraph. A number of closure equations are also
needed. The gas density is calculated using the ideal gas
law. The particle-particle interaction force is assumed to be
AIChE Journal
December 2005
x j k kij
(8)
kij k
1 U kl
U kl
2 k S kij
x l ij
3 x l ij
(9)
1 U ki U kj
2 xj
xi
(10)
3103
U
) gug i pup i
Pij K (U
gi
pi
xj
g
p
U
) pup i gug i
Pij K (U
pi
gi
xj
p
g
U pupi 0
x i p pi
(11)
k k k
k
k k
gg ug i ug j g g gi
xj g gij
U
U
p pU
pi
t
x j p p pi pj
p
kuki tk
(14)
k k
0
U
t
x i k k gi
(18)
2
U
gi
gj
ij k
xj
xi
3
(19)
1
u u
2 gi gi
(20)
tg C
k2
(21)
up i up j tp
2
U
U
pi
pj
pl
ij kp tp
xj
xi
3
xl
(22)
(15)
1
u u
2 pi pi
(23)
U
U
g gU
gi
t
x j g g gi gj
3104
k
xi
0
U
x i p pi
(17)
ug i ug j tg
(13)
k
kk
k
k
k
(16)
pp up i up j p p gi
xj p pij
(12)
December 2005
AIChE Journal
p p
.
K
18
g gP k P kp p
tg ( g g)
g g
g g
U gj
t
xj
xj
xj
g g
(C P C 2) e p
k 1 k
AIChE Journal
U
gi
xj
ugiupi 2k
(32)
(28)
kp k
(33)
TL
m TL
(34)
(35)
(29)
December 2005
TL
m TL
(27)
P k ug i ug j
K
TL
m
k T L
g
m T L ( m T L) 2 x j x j
(26)
Turbulence model
(31)
(25)
tg ( g gk)
g gk
g gk
U gj
t
xj
xj k
xj
(30)
T L 0.41
K
ugi upi
g
xj xj
(24)
m
TL
Comparing the drag term in the model by 18 with the drag term
in Eq. 2, the particle response time can be written
K
1
U
)
ugiugi ugiupi
u (U
pi
g
p p gi gi
ep
tp tg
K 2
U gi U pi
g
(36)
3105
Numerical method
The densities of both phases are assumed to be constant and
a solver for incompressible two-phase ow is used. A thorough
description of the method is given by 38. The transport equations for the turbulence are solved explicitly for each time step
from the latest calculated velocity led. From the k and elds,
new turbulent viscosities for the two phases, and the turbulent
kinetic energy for the particles phase are obtained. These are
then used in the momentum equations for the next time step.
The obtained solution will not completely stationary. The particles will buildup in the channel, and then be swept away from
time to time. This will give a solution that will vary to a small
extent in time, but the variations will be small. This is a true
physical effect if the velocity of the gas is not high enough to
keep all particles suspended. By integrating the velocity and
concentration prole time-averaged proles comparable to the
measured proles are obtained.
is applied to the particle phase. Here the slip parameter suggested by 39 is used
Calculations
The computational domain is a two-dimensional channel
with a width of 0.1 m, and a length of 9.0 m. A coordinate
system is chosen that has the x 1 axis along the length of the
channel and the x 2 axis across the channel and directed upwards. The origin of the coordinate system is located at the
lower wall at the inlet. The grid has 32 cells across the channel
and 900 along the channel length. The grid is rened at the
walls, and the wall cell is 0.77 mm in the x 2 direction, which
is at y between 30 and 40 for the present ow cases.
The horizontal walls are impermeable and a noslip condition
is used for the gas phase. A partial slip condition given by
U p1 L
U p1
x2
(37)
d p
6 2 p
(38)
December 2005
AIChE Journal
Conclusions
The agreement between model predictions and measurements is good. Further investigations should include a 3-D
simulation of a horizontal ow.
Acknowledgments
The authors would like to thank the Swedish Foundation for Strategic
Research (SSF) and SCA Research AB for nancial support of this work.
Notation
C1
C2
C sp
d
C
d
C fib
d
C ki
C
D
df
dp
ep
F dr
i
dr
F
dr
F sp
f
K
k
kp
L
d
M kIi
m
P
Pk
P kp
December 2005
TL
t
U gi
in
U
U ki
U pi
constant in k -model
constant in k -model
drag function for spherical particle
drag function for straight ber perpendicular to the velocity
drag function for ber
particle-particle force, N/m3
constant in k -model
dia. of test pipe, m
mean dia. of the bers, m
mean dia. of the spherical particles, m
additional term in the equation due to the particle presence, m2/s4
drag force on a single particle, N
drag force on ber perpendicular to the velocity, N
drag force on spherical particle, N
function in wake production model, 1/s
drag function, kg/m3 s
turbulent kinetic energy in the gas phase, m2/s2
turbulent kinetic energy in the particle phase, m2/s2
slip parameter according to 39, m
generalized drag, N/m3
loading ratio
pressure, N/m2
production of gas-phase turbulent energy by the mean velocity
eld, m2/s3
additional production of gas-phase turbulent energy due to the
particles, m2/s3
Lagrangian integral time scale, s
time, s
mean gas velocity in the i direction, m/s
inlet mean velocity, m/s
mean velocity of phase k in the i direction, m/s
mean gas velocity in the i direction, m/s
3107
Ur
ugi
upi
V ki
Vp
xi
y
p
p
1
2
ij
tg
tp
k
g
p
d1
d2
k
kij
m
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
Literature Cited
1. Thomas AD. Predicting the deposit velocity for horizontal turbulent
pipe ow of slurries. Int J Multiphase Flow. 1979;5:113129.
2. Oroskar AR, Turian RM. The critical velocity in pipeline ow of
slurries. AIChE J. 1980;26:550 558.
3. Davies JT. Calculation of critical velocities to maintain solids in
suspension in horizontal pipes. Chem Eng Sci. 1987;42:16671670.
4. Cabrejos FJ, Klinzing GE. Pickup and saltation mechanisms of solid
particles in horizontal pneumatic transport. Powder Technol. 1994;79:
173186.
5. Tsuji Y, Oshima T, Morikawa Y. Numerical simulation of pneumatic
conveying in a horizontal pipe. KONA. 1985;3:38 51.
6. Huber N, Sommerfeld M. Modelling and numerical calculation of
dilute-phase pneumatic conveying in pipe systems. Powder Technol.
1998;99:90 101.
. Gas-solid ow with
7. Hussainov M, Kartushinsky A, Mulgi A, Rudi U
a slip velocity of particles in a horizontal channel. J Aerosol Sci.
1996;27:4159.
8. Levy A, Mooney T, Marjanovic P, Mason DJ. A comparison of
analytical and numerical models with experimental data for gas-solid
ow through a straight pipe at different inclinations. Powder Technol.
1997;93:253260.
9. Tsuji Y, Morikawa Y. LDV measurements of an aid-solid two-phase
ow in a horizontal pipe. J Fluid Mech. 1982;120:385 409.
10. Rashidi M, Hetsroni G, Banerjee S. Particle-turbulence interaction in
a boundary layer. Int J Multiphase Flow. 1990;16:935949.
11. Kulick JD, Fessler JR, Eaton JK. Particle response and turbulence
modication in fully developed channel ow. J Fluid Mech. 1994;227:
109 134.
12. Kaftori D, Hetsroni G, Banerjee S. The effect of particles on wall
turbulence. Int J Multiphase Flow. 1998;24:359 386.
13. Tsuji Y, Morikawa Y, Shiomi H. LDV measurements of an air-solid
two-phase ow in a vertical pipe. J Fluid Mech. 1984;139:417 434.
14. Sato Y, Hanzawa K, Maeda M. Interactions between particle wake and
3108
December 2005
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
Manuscript received Apr. 28, 2004, and revision received Apr. 19, 2005.
AIChE Journal
SEPARATIONS
D. H. Reneker
Maurice Morton Institute of Polymer Science, University of Akron, Akron OH 44325
DOI 10.1002/aic.10564
Published online August 10, 2005 in Wiley InterScience (www.interscience.wiley.com).
A number of factors inuence the efciency and the economics of the separation of
dispersed liquid drops in an immiscible liquidliquid mixture. One important factor that
controls the performance of a lter medium in the separation is the ber size. Electrospun
polymer nanobers have diameters in the nanometer range and are arbitrarily long. The
experimental results in this work show that adding nanobers to conventional micronsized brous lter media improves the separation efciency of the lter media but also
increases the pressure drop. An optimum in the performance occurs (signicant increase
in efciency with minimal increase in pressure drop) with the addition of about 1.6% by
mass of 250 nm diameter nylon 6 nanobers to 5 micron diameter glass bers in the lters.
2005 American Institute of Chemical Engineers AIChE J, 51: 3109 3113, 2005
Introduction
Glass-ber lter media augmented with small amounts of
nanobers show improvement of separation efciency. Filter
media with equal amounts of nanobers made of MPD-I [metaaramid, poly(meta-phenyleneisophthalamide)], nylon 6, or
polyacrylonitrile (PAN) were tested for liquidliquid coalescence ltration in a previous article.1 The aim of the study
reported in this article is to present the effects of varying the
amount of nylon 6 nanober added to the glass-ber media on
the pressure drop and the separation efciency in liquidliquid
coalescence lter performance. The nanobers are produced by
electrospinning and have diameters typically in the range of
10 500 nm.2-5
The electrospinning process is driven by the electrical forces
on free charges on the surface or inside of a polymeric liquid.
When the free charges, generally ions, in the polymer solution
move in response to the electric eld, they quickly transfer a
Correspondence concerning this article should be addressed to G. G. Chase at
gchase@uakron.edu.
AIChE Journal
December 2005
3109
December 2005
(1)
n O x
n I x
(2)
AIChE Journal
Filter
Glass-Fiber
Mass (g)
Binder*
Mass
(g)
Nanober
Mass (g)
Total Filter
Sample
Mass (g)
A
B
C
D
E
0.500
0.500
0.500
0.500
0.500
0.325
0.334
0.361
0.359
0.374
0
0.007
0.014
0.020
0.039
0.825
0.841
0.875
0.879
0.913
*Calculated by subtracting the ber amounts from the total lter mass.
December 2005
drop size on the horizontal axis is the average drop size of the
particular bin (the bin sizes in the BR8 are specied by the
largest particle size in that particular bin). The size distribution
marked as Inlet is the size distribution measured from the
sample point upstream of the lter. The distribution marked as
No-Filter was measured at the outlet stream sample point
between the settling tank and the reservoir when no lter was
in the holder. The difference between the Inlet and NoFilter curves shows some of the particles are separated out by
the settling tank even when there is no lter in the line. The
distribution curves marked as Filter A through Filter E
were similarly measured at the outlet stream sample point and
are the steady-state downstream distributions when the lters
are placed in the line. With lters in the line the downstream
numbers of particle concentrations are signicantly reduced.
The design of the settling tank may inuence these results;
thus, the same settling tank was used for all experiments and
these results are comparable between the experiments for relative performance.
Figure 5 shows the separation efciency distributions. The
system with no lter approaches 100% efciency for drop
sizes 125 microns. All of the lters approach 100% efciency for sizes 100 microns. Filter A shows a separation
efciency of 67% for drop size of 125 microns, but this low
efciency is attributed to one drop detected in the outlet stream
from Filter A, whereas the inlet stream contained three such
drops.
In the 0 to 100 micron drop size range signicant differences in performance are noted. Filter A performs better
than No Filter; Filter B performs better than Filter A; and
Filters C, D, and E all perform about the same, all better than
Filter B.
The steady-state pressure drops across the lters are shown
in Figure 6. The lters supplemented with nanobers have a
higher pressure drop than that of lters without nanobers. The
change in pressure drop with the addition of nanobers is
almost linear. Because Filters C, D, and E have similar high
capture efciencies then an optimum design is one that has the
Vol. 51, No. 12
3111
least pressure drop among the lters with the highest capture
efciencies. This means that, of the lters tested here, Filter C
with 0.014 g of nanober (about 1.6% of the lter mass) is the
optimum design.
The pore sizes in the microber lters (no nanobers) are
typically the same order of magnitude as that of the ber
diameter (about 5 microns). Large droplets are easily captured
by interception with multiple bers and the efciency for
microber lters is high for the larger drops. A greater fraction
of smaller drops near the pore size may pass through the pores
without coalescing into larger drops. Thus the lter efciency
decreases for smaller droplet sizes.
Results of this study conrm that the addition of nanobers
improves the lter performance. The nanobers crisscross over
pore spaces, making the pore openings much smaller. Thus it is
more difcult for smaller drops to pass through the pores
without coalescing with other drops.
For the drop sizes considered herein the dominant mechanism of capture is by direct interception and inertial impaction
with multiple bers. Most of the droplets are larger than the
pore openings, and thus the mechanism of single-ber capture
does not apply. The small diameter of the nanober allows us
to reduce the pore sizes with the addition of only small amounts
of nanobers and yet maintain a high porosity.
Other mechanisms, such as surface wetting properties, may
inuence the coalescence performance. We consider the interception of the drops by the nanobers to be more important in
this work.
Conclusions
In this work, the effect of different amounts of polymer
nanobers added to the lter media is experimentally evaluated. The addition of nanobers to glass-ber lter media
improves the capture efciency, but also causes an increase in
pressure drop. The experimental results show that an optimal
amount of nanober enhances the capture efciency but does
not cause excessive pressure drop. For the materials tested
herein 1.6% nanober by mass is optimal.
3112
December 2005
Acknowledgments
This work was supported by Ahlstrom Paper Group, Donaldson Company, Fleetguard Inc., Hollingsworth and Vose, Parker Hannin, and
Produced Water Seminar. This work was also partially supported by
National Science Foundation Grants CTS-0310429 and DMI-0403835.
Literature Cited
1. Shin C, Chase GG. Water-in-oil coalescence using glass ber lters
augmented with polymer nanobers. AIChE J. 2004;50:343-350.
2. Reneker DH, Chun I. Nanometre diameter bres of polymer, produced
by electrospinning. Nanotechnology. 1996;7:216-223.
3. Reneker DH, Yarin AL, Fong H, Koombhongse S. Bending instability
of electrically charged liquid jets of polymer solutions in electrospinning. J Appl Phys. 2000;87:4531-4547.
4. Shin YM, Hohman MM, Brenner MP, Rutledge GC. Electrospinning:
A whipping uid jet generates submicron polymer bers. Appl Phys
Lett. 2001;78:1149-1151.
5. Gibson P, Schreuder-Gibson H, Rivin D. Transport properties of
porous membranes based on electrospun nanobers. Colloids Surf A.
2001;187/188:469-481.
6. Doshi J, Reneker DH. Electrospinning process and applications of
electrospun bers. J Electrostat. 1995;35:151-160.
7. Deitzel JM, Kleinmeyer J, Harris D, Beck Tan NC. The effect of
processing variables on the morphology of electrospun nanobers and
textiles. Polymer. 2001;42:261-272.
8. Sareen SS, Rose PM, Gudesen RC, Kintner RC. Coalescence in brous
beds. AIChE J. 1966;12:1045-1050.
9. Fahim MA, Akbar AM. Removal of ne oily hazes from wastewater
using deep brous bed coalescer. J Environ Sci Health. 1984;A19:
299-319.
10. Othman FM, Fahim MA, Jeffreys GV, Mumford GJ. Prediction of
predominant mechanisms in the separation of secondary dispersion in
a brous bed. J Dispers Sci Technol. 1988;9:91-113.
11. Secerov Sokolovic RM, Sokolovic SM, -Okovic BD. Effect of
working conditions on bed coalescence of an oil-in-water emulsion
using a polyurethane foam bed. Ind Eng Chem Res. 1997;36:49494953.
12. Moses SF, Ng KM. A visual study of the breakdown of emulsions in
porous coalescers. Chem Eng Sci. 1985;40:2339-2350.
13. Hajra MG, Mehta K, Chase GG. Humidity, temperature, and polymer
nanobers on drop coalescence in glass ber media. Sep Purif Technol.
2003;30:79-88.
14. Hazlett RN. Fibrous bed coalescence of water: Role of a sulfonate
surfactant in the coalescence process. Ind Eng Chem Fundam. 1969;
8:633-640.
15. Hazlett RN. Fibrous bed coalescence of water: Steps in the coalescence process. Ind Eng Chem Fundam. 1969;8:625-632.
16. Akbar AM, Othman FM. Prediction of oil saturation in a brous bed
coalescer from pressure drop data. J Dispers Sci Technol. 1989;10:
697-713.
17. Chase GG, Beniwal V, Venkataraman C. Measurement of uni-axial
ber angle in non-woven brous media. Chem Eng Sci. 2000;55:21512160.
AIChE Journal
AIChE Journal
December 2005
Manuscript received Feb. 20, 2004, and revision received Apr. 5, 2005.
3113
The solvent is the core of extractive distillation, and a suitable solvent plays an
important role in the economical design of extractive distillation. Computer-aided molecular design (CAMD) has been applied to rapidly screen the solvents for separating
hydrocarbons by extractive distillation. The systems of propane/propylene, n-butane/1butene, and n-heptane/benzene, respectively, as the representatives of C3, C4, and C6
hydrocarbons were investigated, and the potential solvents were selected by means of
CAMD. The designed results were further proven by experiments and process simulation.
The mechanism for separating hydrocarbons by extractive distillation is based on the
different uidities of the electron cloud of COC (no double bond), CAC (one double
bond), and ACH (aromatic carbon ring) bonds and thus different interactions between
solvent and hydrocarbon molecules. To improve the separation ability of the main solvent,
one strategy is to add some additive that can form hydrogen bonding with the main solvent
to make into a mixture. 2005 American Institute of Chemical Engineers AIChE J, 51:
3114 3121, 2005
Introduction
Extractive distillation is commonly applied for separating
hydrocarbons with close boiling points, such as C4, C5, and C6
mixtures and so on.1-3 In extractive distillation, an additional
solvent (entrainer or separating agent) is used to alter the
relative volatility of the components to be separated. In this
way, it is possible to obtain one pure component at the top of
one column and the other, together with the solvent at the
bottom, which may then be easily separated in a secondary
distillation column because of the high boiling point of the
solvent. The solvent does not need to be vaporized in the
extractive distillation process.
It is known that the solvent is the core of extractive distillation, and a suitable solvent plays an important role in the
Correspondence concerning this article should be addressed to B. Chen at
chenbh@mail.buct.edu.cn or Z. Lei at leizhg@mail.buct.edu.cn.
3114
December 2005
AIChE Journal
CAMD Program
CAMD is conducted in the following four steps.
December 2005
3115
Method
ij
i Pi0
j Pj0
i
j
1 MWi
SP
i MWs
Pure component data bank or by adding
group parameters32
Pure component data bank or by adding
group parameters3234
Pure component data bank or Joback
method32,33
Antoine equation or Riedel equation32,33
By drawing the curves of x (mole fraction in
the liquid phase)y (mole fraction in the
vapor phase) to judge
Sij
Case Studies
The systems of propane/propylene, n-butane/1-butene, and
n-heptane/benzene are investigated because they are commonly
encountered in industry and may be separated by extractive
distillation.3
The separation mechanism for separating hydrocarbons by
extractive distillation is based on the different uidities of the
electron cloud of COC (no double bond), CAC (one double
bond), and ACH (aromatic carbon ring) bonds. The greater the
uidity, the easier can the group be polarized by the polar
solvent. The uidity of the electron cloud of CAC (or ACH)
bond is greater than that of the COC bond so that propane (or
3116
December 2005
AIChE Journal
Table 2. Results of CAMD for the Separation of n-Heptane (1) and Benzene (2) at T 303.15 K
Azeotropic Judgment
No.
Molecular Structure
MW
T b (K)
12
S
12
SP
1 and Solvent
2 and Solvent
1
2
3
4
5
6
7
8
9
10
DMSO
CH2CNOCH2CN
CH3COOOCH2OCH2CN
DMF
CH2NH2OCH2CN
NMP
CH2NH2OCH2OCOOH
ACN
CH3COOOCH2OCH2OCH2CN
OHOCH2OCH2CN
78.1
80.0
113.0
73.1
70.0
99.1
89.0
41.0
127.0
71.0
462.2
495.3
474.3
426.15
442.9
477.2
486.3
354.8
497.2
462.5
9.782
8.759
7.026
5.767
5.604
5.552
5.242
5.134
5.086
5.067
19.938
17.853
14.321
11.754
11.422
11.317
10.685
10.465
10.366
10.328
0.318
0.312
0.354
0.686
0.468
0.756
0.543
0.621
0.417
0.210
yes
yes
no
yes
yes
no
no
yes
no
yes
no
no
no
no
no
no
no
yes
no
no
December 2005
3117
Table 3. Results of CAMD for the Separation of Propane (1) and Propylene (2) at T 303.15 K
Azeotropic Judgment
No.
Molecular Structure
MW
Single solvent
80.0
1
CH2CNOCH2CN
2
ACN
41.0
3
CH3COOOCH2OCH2NH2
103.0
113.0
4
CH3COOOCH2OCH2CN
5
NMP
99.1
6
CH3COOCH2OCH2NH2
87.0
97.0
7
CH3COOCH2OCH2CN
8
CH3COOOCH2OCH2OCH2CN
127.0
9
CH3COOOCH2OCH2OCH2NH2
117.0
10
CH3COOCH2OCH2OCH2NH2
101.0
70.0
11
CH2NH2OCH2CN
12
CH3COOCH2OCH2OCOOCH3
130.0
13
CH3COOOCH2OCH2OCOOCH3
146.0
111.0
14
CH3COOCH2OCH2OCH2CN
Additive added to ACN to make into a mixture (10 wt %)
15
H2O
18.0
Additive added to NMP to make into a mixture (10 wt %)
16
H2O
18.0
Tb
(K)
12
S
12
SP
1 and
Solvent
2 and
Solvent
495.3
354.8
421.9
474.3
477.2
417.5
470.0
497.2
444.8
440.4
442.9
449.0
453.3
492.8
1.802
1.704 (1.620)
1.634
1.629
1.620
1.576
1.542
1.523
1.505
1.486
1.484
1.480
1.468
1.458
2.172
2.055
1.970
1.964
1.954
1.900
1.859
1.836
1.815
1.791
1.790
1.784
1.770
1.758
0.142
0.267
0.451
0.149
0.241
0.406
0.170
0.181
0.456
0.420
0.187
0.190
0.182
0.197
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
373.2
1.846 (1.750)
2.226
0.248
no
no
373.2
1.969
2.374
0.175
no
no
Process simulations of the separation of propane and propylene by ordinary distillation and by extractive distillation
with the mixture of ACN and water as the solvent were respectively made with PRO/II software. The ordinary distillation column was operated at pressure 1910 kPa and reux ratio
15.0 with 120 theoretical plates. The extractive distillation
process consisted of four sections: extractive distillation column with 50 theoretical plates; stripping column with 10 theoretical plates; water scrubber column with 10 theoretical
plates; and solvent recovery column with 20 theoretical plates.
In the case of the feeding composition of propylene 82.0 wt %,
feed rate 32,000 kg/h, production rate 26,500 kg/h, and molar
composition of product propylene 99.3%, the extractive
distillation process saved 13.2% reboiler load, 79.1% condenser load, and 25% theoretical plates when compared to
ordinary distillation process.
December 2005
Thermodynamic Analysis
By comparison of Tables 3 and 4, it is found that from
numbers 1 to 13 the solvents screened by CAMD and their
order in both systems is consistent. The reason may be explained by the similar molecular structure of C3 and C4 hydrocarbons (a difference of only one carbon chain) and the
same separation mechanism.
The experimental results obtained by inert gas stripping and
gas chromatography are also given in parentheses in Table 4.
However, it is found that the mixture of ACN and ethylenediVol. 51, No. 12
AIChE Journal
Table 4. Results of CAMD for the Separation of Butane (1) and Butene (2) at T 303.15 K
Azeotropic Judgment
No.
Molecular Structure
Tb
(K)
12
S
12
SP
1 and
Solvent
2 and
Solvent
495.3
354.8
421.9
474.3
477.2
417.5
470.0
497.2
444.8
440.4
442.9
449.0
453.3
1.780
1.684 (1.696)
1.614
1.609
1.601
1.557
1.524
1.505
1.487
1.468
1.466
1.462
1.451
2.172
2.055
1.970
1.964
1.954
1.901
1.859
1.836
1.815
1.791
1.790
1.784
1.770
0.111
0.238
0.477
0.122
0.225
0.414
0.143
0.157
0.491
0.437
0.161
0.168
0.161
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
373.2
495.3
462.5
399.2
352.8
390.4
337.8
443.5
351.5
1.824 (1.880)
1.692
1.649
1.640 (1.725)
1.631 (1.707)
1.626 (1.787)
1.610
1.607
1.579 (1.713)
2.226
2.065
2.012
2.002
1.990
1.984
1.965
1.961
1.927
0.209
0.223
0.219
0.245
0.251
0.238
0.242
0.225
0.250
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
373.2
1.945
2.374
0.139
no
no
MW
Single solvent
80.0
1
CH2CNOCH2CN
2
ACN
41.0
3
CH3COOOCH2OCH2NH2
103.0
113.0
4
CH3COOOCH2OCH2CN
5
NMP
99.1
6
CH3COOCH2OCH2NH2
87.0
97.0
7
CH3COOCH2OCH2CN
8
CH3COOOCH2OCH2OCH2CN
127.0
9
CH3COOOCH2OCH2OCH2NH2
117.0
10
CH3COOCH2OCH2OCH2NH2
101.0
70.0
11
CH2NH2OCH2CN
12
CH3COOCH2OCH2OCOOCH3
130.0
13
CH3COOOCH2OCH2OCOOCH3
146.0
Additive added to ACN to make into a mixture (10 wt %)
18.0
14
H2O
15
CH2CNOCH2CN
80.0
16
OHOCH2OCH2CN
71.0
17
CH3OCH2OCH2OCH2OCH3COO
116.0
18
CH3COOCH2OCH3
72.0
19
CH2NH2OCH2NH2
60.0
32.0
20
CH3OH
21
OHOCH2OCH2NH2
61.0
46.0
22
CH3OCH2OOH
Additive added to NMP to make into a mixture (10 wt %)
23
H2O
18.0
amine also can give rise to higher relative volatility than single
ACN or single ethylenediamine.
Why do the mixtures of ACN and water and that of ACN and
ethylenediamine show good separation ability? This may be
explained from Prausnitz and Andersons solution thermodynamics.36,37
For the hydrocarbons to be separated by extractive distillation, selectivity (which is consistent with relative volatility) is
related to the various energy terms leading to the desired
nonideality of solution, which is the basis of extractive distillation, and can be expressed as
is, the polar effect, the dispersion effect, and the inductive
effect of the solvent. It is convenient to rewrite as
RT ln S23 P D I
(2)
(3)
RT ln S23 V2 V3 V2 1n 2 V3 1n 3
2
1p
AIChE Journal
Case 1
Case 2
Case 3
2994
112
6
2994
112
12
2994
112
11
December 2005
3119
Figure 2. Inuence of carbon number on relative volatility of n-butane to 1-butene at innite dilution
for the solvents whose molecular structures
are ROC'N: R CnH2[in]n1 (n carbon number 1, 2, 3, 4, 5, 6).
Conclusions
CAMD has been programmed to screen the solvents for
separating hydrocarbons by extractive distillation. The systems
of propane/propylene, n-butane/1-butene, and n-heptane/benzene are investigated because they are commonly met in industry and may be separated by extractive distillation. By
means of CAMD, it is found that for the separation of propane
and propylene, the nal potential solvent is ACN (no hydrolysis) or the mixture of ACN and water (hydrolysis); for the
separation of n-butane and 1-butene, the nal potential solvent
is ACN (no hydrolysis) or the mixture of ACN and water
(hydrolysis), although the mixture of ACN and ethylenediamine (no hydrolysis) also shows high separation ability; for the
separation of n-heptane and benzene, the nal potential solvent
is NMP.
3120
December 2005
AIChE Journal
strategy is to add some additive that can form hydrogen bonding with the main solvent to make into a mixture.
In this work the solvents of extractive distillation are restricted only to liquid solvents. In fact, so far they have been
developed to include inorganic salts, polymers, and ionic liquids. Unfortunately, the group parameters of polymer and ionic
liquid are vacant in the parameter table except that there are a
few group parameters for inorganic salts. Even for liquid
solvents, there are still about 54% of the UNIFAC group
interaction parameters missing because experimental data are
needed to ll them. This is where CAMD is limited. For the
development of calculation techniques, it is thought that, besides the efforts to blend the computation methods of target
properties with the screening of potential solvents and replenish the group parameters, were CAMD combined with other
software (such as Excel, PRO/II, ASPEN PLUS), its functions
would become stronger and the designed results would be more
reliable.
15.
16.
17.
18.
19.
20.
21.
22.
23.
Acknowledgments
This work is nancially supported by the National Nature Science
Foundation of China under Grant 20406001.
24.
Literature Cited
25.
1. Berg L. Separation of benzene and toluene from close boiling nonaromatics by extractive distillation. AIChE J. 1983;29:961-966.
2. Chen B, Lei Z, Li J. Separation on aromatics and non-aromatics by
extractive distillation with NMP. J Chem Eng Jpn. 2003;36:20-24.
3. Lei Z, Li C, Chen B. Extractive distillation: A review. Sep Purif Rev.
2003;32:121-213.
4. Joback KG, Stephanopoulos G. Searching spaces of discrete solutions:
The design of molecules possessing desired physical properties. Adv
Chem Eng. 1995;21:257-311.
5. Venkatasubramanian V, Chan K, Caruthers JM. Genetic algorithmic
approach for computer-aided molecular design. ACS Symp Ser. 1995;
589:396-414.
6. Churi N, Achenie LEK. Novel mathematical programming model for
computer aided molecular design. Ind Eng Chem Res. 1996;53:37883794.
7. Marcoulaki EC, Kokossis AC. Molecular design synthesis using stochastic optimisation as a tool for scoping and screening. Comput Chem
Eng. 1998;22:S11-S18.
8. Mavrovouniotis ML. Design of chemical compounds. Comput Chem
Eng. 1998;22:713-715.
9. Meniai AH, Newsham DMT, Khalfaoui B. Solvent design for liquid
extraction using calculated molecular interaction parameters. Chem
Eng Res Des. 1998;76(A11):942-950.
10. Ourique JE, Telles AS. Computer-aided molecular design with simulated annealing and molecular graphs. Comput Chem Eng. 1998;22:
S615-S618.
11. Pistikopoulos EN, Stefanis SK. Optimal solvent design for environmental impact minimization. Comput Chem Eng. 1998;22:717-733.
12. Raman VS, Maranas CD. Optimization in product design with properties correlated with topological indices. Comput Chem Eng. 1998;
22:747-763.
13. Harper PM, Gani R, Ishikawa T, Kolar P. Computer aided molecular
design with combined molecular modeling and group contribution.
Fluid Phase Equilib. 1999;158-160:337-347.
14. Hostrup M, Harper PM, Gani R. Design of environmentally benign
AIChE Journal
December 2005
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
Manuscript received Dec. 3, 2004, and revision received Mar. 27, 2005.
3121
Krzysztof Kaczmarski
Faculty of Chemistry, Rzeszow University of Technology, 35-959, Rzeszow, Poland
DOI 10.1002/aic.10565
Published online September 2, 2005 in Wiley InterScience (www.interscience.wiley.com).
Introduction
Liquid chromatography has become an important process for
the extraction and/or the purication of drug intermediates in
the pharmaceutical industry. It is of special interest for fragile
molecules, for those that are obtained as components of very
complex mixtures of similar compounds, and for difcult separations. Typical examples are the preparation of peptides,
proteins, and enantiomers. In all these instances, the interest for
the modeling of specic HPLC separations has much increased
during the last few years because computer-assisted modeling
of processes permits a considerable reduction in the time and
manpower needed to start a process or to start up a unit after
Correspondence concerning this article should be addressed to G. A. Gulochon at
guiochon@utk.edu.
3122
December 2005
AIChE Journal
increasing cost in time spent and chemicals consumed.4 Current research aims at reducing these costs but will not change
the need for these data. If the relationship between equilibrium
isotherms, the calculation of elution band proles, and the
modeling of process performance is clearly established, the
inuence of the mass transfer kinetics on band proles and
process performance remains insufciently understood.
For its better understanding, the kinetics of equilibration
between the uid and the solid phases in chromatography is
divided into several steps that can be grouped into two broad
classes, axial dispersion and the mass-transfer resistances. The
former class involves axial diffusion, the consequence through
Fick laws of the concentration gradient that migrates along the
column, and eddy diffusion, which arises from the ow pattern.
The chromatographic columns are operated under a rather slow
stream of the uid phase. In liquid chromatography, the Reynolds numbers of these ows are typically between 0.01 and
0.001, ensuring creeping ow conditions.5 There are no significant eddies but the distribution of the average residence times
along the different stream-lines has a signicant width, causing
eddy diffusion.2,5 The contributions of axial dispersion to band
broadening tend to decrease with increasing ow velocity.
By contrast, the effects of the mass-transfer coefcients
increase with increasing mobile phase velocity. They include
mass transfer between the uid phase stream percolating
through the bed and the uid stagnant inside the particles (or in
the porons, in the case of monolithic columns), diffusion
through this stagnant uid across the particles, and the kinetics
of adsorption desorption onto the solid surface. The rst two
of these contributions are called the external and the internal
mass transfer resistances, respectively. The latter contribution
has been abundantly studied and our recent results have demonstrated the importance of surface diffusion in reversed-phase
liquid chromatography.6 The external mass transfer resistance
has been much less investigated. It is usually estimated in
kinetic studies from the Wilson and Geankoplis correlation.7
This remarkable equation was derived, however, from measurements carried out on a column packed with 1/4 solid beads
of benzoic acid percolated with a stream of water. The validity
of this equation for columns packed with silica particles that
are a thousand times smaller and exchange solutes with the
external uid through a multitude of tiny pores has never been
demonstrated.
The goal of this work was to investigate the external mass
transfer kinetics by determining, under linear conditions, its
rate constant for a series of compounds on columns packed
with the same packing material, but using lots having average
particle sizes spanning a wide range, between 3 and 50 m,
hence, under experimental conditions spanning a 100-fold
range of the Fick number. The dependence of the external rate
constant on the particle size should provide new clues on the
mass transfer-kinetics in chromatography.
Theory
General rate model (GR)
Any serious investigation of the inuence of the masstransfer kinetics on the band proles in chromatography must
use the general rate model of chromatography. Although there
are no algebraic, only numerical solutions for this model, a
solution is available for the Laplace transform in the linear
AIChE Journal
December 2005
C
2C
C
u
eD L
t
x
x2
1 3 k
e,i
C C p,i r R i
Ri
ext
(1)
q i
C p,i
1
C p,i
r2
1 p
D eff 2
t
t
r r
r
(2)
Vi
Va
(3)
1
Vol. 51, No. 12
L
Cttdt
0
Ctdt
u0
(4)
3123
Ctt 1 2dt
Ctdt
(5)
H HL Hk HD
2u 01 e P 1 P K 2 R
R2
2
3k ext 15D eff
0
(6)
1 e P 1 P K 2
02
eD L 02
u 02
(7)
H0
R2
f d
R
A
2
3k ext 15D eff
0
f 1 eR/ 2k ext p 1 p K 2
(8)
d 1 eR 2/15D eff P 1 p K 2
(9)
0 e 1 e P 1 P K
(10)
12
2
(11)
L L 2
2
N
1
(12)
2u 0 e
L f d
02
(15)
(16)
(17)
p
o e
1 e1 p 1 p
(18)
(13)
eD L
2 L
2 H0
12 2u 0
u0
(14)
2L
f d 2 inj 2 sys
u0 L
3124
2 eD L
u0
Re ud/
(20)
Sc / D m
(21)
AIChE Journal
Sh k extd/D m
From the correlation Eq. 19, it follows that the external mass
transfer rate constant is inversely proportional to the particle
diameter to the power (2/3).
k ext
1.09u 0 1/3 D m
e
d
2/3
(23)
AM s 0.5
V 0.6
A
(24)
2
e
eD L
B1 2 2
du 0
2
Re Sc
(25)
with
B
Re Sc2
Re Sc
1 p2
4 1 e
1614 1 e 2
2
1
p1 p3 exp
4 1 e
1
p1 pRe Sc
2
1
(28)
Particle-size distribution
A relationship can be derived easily between the rst two
moments of the elution bands, the column HETP under linear
conditions, and the parameters of the general rate model if the
column bed is assumed to consist in spherical particles having
all the same diameters. However, all actual packing materials
have a particle-size distribution (PSD) of nite width. Typical
examples of the PSD of contemporary packing materials and
adsorbent particles are discussed by Neue.17 For practical calculations, the PSD is replaced by the mean-particle diameter.
There exist several methods to calculate the mean diameter of
a particulate material, but for the description of the mass
transfer processes, such as those that take place in chromatography, the volume moment mean diameter, noted d 43 , is recommended17
d 43
i n id i4
i n id i3
(29)
Experimental
Equipment
(26)
(22)
December 2005
An HP 1090 (Hewlett-Packard, Palo Alto, CA) liquid chromatograph was employed for all the experimental determinations. The instrument is equipped with a multisolvent delivery
system, an automatic sample injector with a 100-L loop, a
diode array UV-detector and a computer data station that
controls its operation. The signal of the UV detector was
acquired at a wavelength of 254 nm.
3125
Description
3 [m]
5 [m]
10 [m]
15 [m]
50 [m]
3.74
10%
1.57
103
383
17.35
3.48
0.603
0.362
0.378
5.21
6%
1.49
98
398
18.13
3.57
0.592
0.369
0.353
10.6
6%
1.59
101
386
17.07
3.03
0.61
0.371
0.38
14.92
7.5%
1.54
107
388
15.56
2.96
0.642
0.379
0.423
47.1
1.82
0.612
0.37
0.384
Ratio of the diameters below which 90% and 10% of the particles, respectively, are found.
Values measured in our laboratory.
Chromatographic conditions
Sets of identical experiments were carried out under different ow rates on a series of stainless steel columns (all 150
4.6 [mm]). Five columns were packed by the manufacturer
with Luna C18 (Phenomenex, Torrance, CA) in different average particle sizes, 3, 5, 10, 15, and 50 [m]. This packing
material is made of spherical particles of porous silica, chemically bonded with octadecylsilane. Their chemical properties
are broadly similar as far as alkylaromatics are concerned. The
details of the column parameters, as supplied by the manufacturers, are shown in Table 1. These columns are similar to one
that was investigated by Kele18 but they belong to more recent
lots and most of them have different particle sizes. The column
temperature was kept constant at 25.0C.
To achieve accurate results, the column was equilibrated for
at least three hours prior to any measurement when the mobile
phase composition was changed. The run-to-run reproducibility of the instrument was characterized by a RSD of the peak
area which was better than 0.50% (six replicates) for an injection volume of 5 L. The rst and second moments of the
elution bands of the solutes were measured for the following
mobile phase ow rates: 0.8, 1.0, 1.2, 1.5, and 1.8 [mL/min] for
average particle sizes of 5, 10, and 15 [m]; 0.6, 1.0, 1.2, and
1.4 [mL/min] for an average particle sizes of 3 [m]; and 1.2,
2.0, 2.4, and 2.8 [mL/min] for an average particle size of 50
[m].
The total porosity t , of each column was derived from the
retention volume of thiourea, that was considered to be unretained on all the columns, an assumption consistent with the
results previously obtained by this group under the experimental conditions used here. The external porosity e , was obtained
from measurements carried out under ISEC mode (see later,
section 3.6). The internal porosity P , was calculated according
to the conventional equation t e (1 e ) p . The
porosities are also listed in Table 1.
December 2005
AIChE Journal
Figure 1. Logarithm of the molecular mass of the samples vs. their retention volumes for a series of
polystyrene standards.
December 2005
Inuence of the Mean Particle Size. The actual mean particle size can differ by as much as 10% from the nominal value,
as seen in Table 1. To gain an accurate knowledge on how the
value of the particle-mean diameter can inuence the second
moment, numerical solutions of the GR model were calculated
for particle diameters equal to 5 and 5.3 [m], on the one hand,
to 50 and 53.5 [m], on the other hand. The mobile phase
velocity was assumed to be equal 1 [ml/min]. The other GR
model parameters (the porosities, the dispersion coefcient,
and the mass-transfer coefcients) were taken from those derived from our experimental data and which are reported later
in this work or can be found in Table 1.
According to the calculated peak proles, the relative difference between the second moment calculated for the two
pairs of actual particle diameters was derived according to the
equation
2,d2 2,d1
2,d1
(30)
2,PSD 2
2,PSD
(31)
and 2 denote
the second moments calculated with the actual PSD, and with
the mean-particle diameter, respectively.
As seen in Table 2, the relative differences between the
correct value of the second moment and the one calculated with
Vol. 51, No. 12
3127
Figure 2. Integral volume fraction f, or ratio of the volume of the particles with a diameter less than
d to the total volume of the particles vs. the
particle diameter.
The volume mean diameter of this material is d 43 12.77
[m].
Figure 3. Comparison between the peak proles calculated by GR for the average particle dia. d43
46.49 [m] (higher peak), and for the true PSD
(smaller peak).
3128
December 2005
0.8
0.071
46.49
1
0.063
1.4
0.056
12.7
1
0.028
5.16
1
0.0098
AIChE Journal
Exp
Eq. 25
5
3.74
4.28 10
5.21
6.30 105
10.6
8.13 105
14.92
1.34 104
47.1
1.17 103
Propylbenzene
b
Exp
5
1.45 10
5.86 105
2.16 105
8.61 105
1.32 104
5.37 104
2.48 104
9.08 104
1.52 103
3.53 103
Eq. 25
5
4.23 10
5.96 105
8.88 105
2.02 104
1.27 103
Butylbenzene
b
Exp
5
1.51 10
6.06 105
2.28 105
9.23 105
1.42 104
5.67 104
2.56 104
9.50 104
1.58 103
3.61 103
Eq. 25b
5
3.53 10
7.15 105
9.29 105
1.46 104
1.28 103
Eq. (28)b
5
1.55 10
6.38 105
2.37 105
9.72 105
1.49 104
5.89 104
2.68 104
9.81 104
1.62 103
3.66 103
1.1 104
2.6 104
2.1 104
4.7 104
4.3 104
9.5 104
6.0 104
1.3 103
1.9 103
3.3 103
Values derived from the slopes of the lines in Figure 4, using Eq. 17.
Values calculated from Eq. 25 or 28.
AIChE Journal
December 2005
3129
3k ext
15
R2
D
1 pRD sK R
R
m p
(32)
(assumed)
D m pR
1 pRD sK R
(33)
2 p 2
p
(34)
p 1.51 p
(35)
d [m]
Experiment
Theoryb
3.74
5.21
10.6
14.92
47.1
0.2
0.162
0.101
0.0806
0.0377
0.070.094
0.0570.082
0.0510.039
0.0310.041
0.0140.019
3130
December 2005
D s [cm2/s]
Fisher
Test
6.34 107
4.926 107
5.04 107
1041
1281
1012
3.45 107
3.44 107
4.23 107
5.57 107
1.44 106
4.59 107
2.45 107
2.28 107
2.19 107
1197
1464
1076
1023
1002
498
3321
6678
1 104
k ext
[cm2/s]
k ext
R p
(36)
AIChE Journal
Dm
(37)
December 2005
2/3
1.06
1.45
1.95
1.03
1.07
1.39
0.46
1.07
1.13
2.02
0.89
1.1
D mi /D mj
E/
P
P/
B
E/
B
b
c
3131
Conclusion
Although the apparent dispersion coefcient and the external
mass-transfer coefcient may depend on the mobile phase
velocity, our experimental results show that this inuence is
smaller than the experimental errors in the range of velocity
investigated. The particle diameter inuences the external mass
transfer coefcient. The experimental results show that the
dependence of the rate constant of the external mass transfer on
the particle diameter is not consistent with the Wilson-Geankoplis correlation. This correlation, often used to account for
HPLC results, is generally accepted as valid in chemical engineering but it was developed for particle sizes nearly 1,000
times larger than commonly used in contemporary chromatography, and for Reynolds numbers that were, mostly, much
larger than those used in liquid chromatography. The range of
validity of the correlation is 0.0015 Re 55 while columns
are run usually at 0.001 Re 0.05, which places them at the
very low end of the recommended range of validity of the
correlation.
Our experimental results suggest, by contrast, that the exterTable 7. Ratios of Effective Surface to Pore Diffusivities and
of Surface to Molecular Diffusivities
3132
December 2005
Solute
D e,s /D p
D s /D m
Ethylbenzene
Propylbenzene
Butylbenzene
0.67
1.12
2.04
4.04 102
3.97 102
4.09 102
AIChE Journal
1
2
viscosity
absolute moment
second central moment
uid density
Subscripts
a denote adsorbent
i i-th fraction of adsorbent particle
s solid phase
Superscript
R means that parameter depends on used particle radius
Acknowledgment
Literature Cited
Notation
A parameter dened by Eq. 15
C concentration in the mobile phase
C p concentration in the stagnant uid phase contained inside pores
d equivalent particle diameter
d 43 volume moment mean diameter
D m diffusion coefcient
D eff effective (or inside-pore) diffusion coefcient
D p pore diffusivity
D L dispersion coefcient
D s surface diffusivity
D s,e effective surface diffusivity
H high equivalent to theoretical plates
H L , H k , H D parameters dened by Eq. 14
H o , H parameters dened by Eqs. 16 and 17
k ext external mass-transfer coefcient
K equilibrium constant
L column length
N number of theoretical plates
M s molecular mass
Pe Peclet number, uL/D L e
q concentration of the solute in the stationary phase
r radial distance from the center of the particle
Re Reynolds number, ud/
R particle radius
Sc Schmidt number, / D m
Sh Sherwood number, k ext d/D m
t time
T absolute temperature
u o average supercial velocity of the mobile phase
V volume
V A molar volume of solute
x longitudinal distance along the column
Greek symbols
A association factor
exponent of the dependence of the rate constant of
external mass transfer on the particle diameter (Eq. 36)
thickness of the thin lm
L , f , d , o parameters dened by Eqs. 710
tortuosity parameter
e external porosity
p internal porosity
t total porosity
AIChE Journal
December 2005
Manuscript received Dec. 20, 2004, and revision received Apr. 7, 2005.
3133
Eric C. Kerrigan
Dept. of Engineering, University of Cambridge, Cambridge CB2 1PZ, U.K.
DOI 10.1002/aic.10626
Published online August 19, 2005 in Wiley InterScience (www.interscience.wiley.com).
The design of a dynamic state feedback receding horizon controller is addressed, which
guarantees robust constraint satisfaction, robust stability and offset-free control of
constrained linear systems in the presence of time-varying setpoints and unmeasured
disturbances. This objective is obtained by rst designing a dynamic linear offset-free
controller and computing an appropriate domain of attraction for this controller. The
linear (unconstrained) controller is then modied by adding a perturbation term, which
is computed by a (constrained) robust receding horizon controller. The receding horizon
controller has the property that its domain of attraction contains that of the linear
controller. In order to ensure robust constraint satisfaction, in addition to offset-free
control, the transient, as well as the limiting behavior of the disturbance and setpoint need
to be taken into account in the design of the receding horizon controller. The fundamental
difference between the results and the existing literature on receding horizon control is
that the transient effect of the disturbance and set point sequences on the so-called target
calculator is explicitly incorporated in the formulation of the receding horizon controller. An example of the control of a continuous stirred-tank reactor is presented. 2005
American Institute of Chemical Engineers AIChE J, 51: 3134 3146, 2005
Introduction
The control of systems in the presence of constraints is an
important task in many application elds because constraints
always arise from physical limitations and quality or safety
reasons. Moreover, in practical applications, disturbances are
usually present and often they are not measurable or predictable. For example, in the chemical industries disturbances arise
from interactions between different plant units, from changes
Correspondence concerning this article should be addressed to G. Pannocchia at
g.pannocchia@ing.unipi.it.
3134
December 2005
AIChE Journal
December 2005
(1a)
zk C zxk
(1b)
(2)
I C A B0 n p
(3)
3135
set point is known but future set point values are unknown. The
set point sequence s takes on values in a polytope p
containing the origin and asymptotically reaches an unknown
steady-state value, that is, s(k) for all k , and there
exists an s such that limk3 s(k) s .
Assumption 3 (Disturbance). At each time instant, current
and future disturbances are unknown. The disturbance sequence d takes on values in a polytope r containing
the origin and asymptotically reaches an unknown steady-state
value d , that is, d(k) for all k and there exists a d
such that limk3 d(k) d .
Remark 2. Note that, unlike many existing results, we do
not assume that the disturbance or set point is constant. Furthermore, unlike,25 we do not assume that the set point and/or
disturbance are generated by a known nite-dimensional exogenous system. The lack of these assumptions in this article
complicates the design of the controller.
Under the earlier assumptions we present a novel method for
designing a dynamic, state feedback receding horizon controller that, for any allowable disturbance and set point sequences
(that is, any innite disturbance and set point sequences that
satisfy Assumptions 2 and 3), accomplishes the goal of asymptotically driving the controlled variable to any given allowable
asymptotic set point, while respecting the state and input constraints, that is
lim zk s
(4a)
k3
xk ,
uk ,
k .
(4b)
min H1 w min Hq wT
w
w
k 1 F k wk
(5)
Augmented system
We will make use of the following auxiliary system in order
to dene the controller dynamics
x k 1 Axk Buk d k xk x k
December 2005
d k 1 d k xk x k
Vol. 51, No. 12
(6a)
(6b)
AIChE Journal
subject to
I C A B0 ux 0I
z
k 1 k uk dk
(7a)
zk k
(7b)
Target calculation
When a nonzero disturbance affects a system (and/or the
current set point s is different from zero), the steady-state
target value of the state and input may need to be shifted in
order to cancel the effect of such a disturbance on the controlled variable.33,34 To this aim, at each sample instant we use
the estimate of the future disturbance and compute the steadystate target ( x , u ) such that one can drive the controlled
variable to the current set point, by solving the following
least-squares problem 34 in which R mm is a positive
denite matrix
x * , s, u * , s : argmin
x ,u
AIChE Journal
1 T
u R u
2
(8a)
December 2005
I I
0
0 0 I s
(8b)
ux **,, ss
x,
u,
x, x,
x,s
u, u, u,s s
(9)
(10)
3137
(11)
then
:
(12)
is strictly stable.
Proof. The statement follows from straightforward block
matrix manipulations,36 where one can show that 2n of the
eigenvalues of are at zero and the rest are the eigenvalues
of A BK.
By dening
: u, Kx, ,
: u,s Kx,s
(13)
(14)
k 1 k dk sk,
(15a)
zk k,
(15b)
where is suitably-dened.
It immediately follows from Lemma 2 that the closed-system
(Eq. 15) is stable if the disturbance and set point are zero (or
asymptotically constant). As a consequence, we introduce the
following standing assumption:
Assumption 4 (Stabilizing gain). The matrix K mn is
chosen such that A BK is strictly stable, is given by Eq.
11, and given by Eq. 13 and : .
The following result states that if the control is given by u
s, then the value of the controlled variable for Eq. 15
is guaranteed to converge to the asymptotic set point s , given
any allowable innite set point and disturbance sequence:
Lemma 3 (Offset-free control). If Assumptions 1 4 hold,
then the solution of the closed-loop system (Eq. 15) satises
limk3 z(k) s for all (0) 3n .
Proof. See the Appendix.
December 2005
the internal model principle37 (roughly) states that a controller that guarantees offset-free control even when the
system parameters are perturbed, must incorporate a model
of the dynamic structure of the disturbance, and the reference signals; in our case, the observer contains a model of
the disturbance and the target calculator contains a model of
the disturbance and set point. This point will be further
illustrated in the example section.
(16)
The maximal constraint-admissible robust positively invariant set for the closed-loop system (Eq. 15) is dened as all
initial states in for which the evolution of the system remains
in for all allowable innite set point and disturbance sequences, that is
: 0 k 1 k dk
sk , sk , dk , k
(17)
AIChE Journal
(18)
(19)
N , s : v mN
AIChE Journal
k 1 k k dk sk
0 , s0 s,
i1 i i di si , i 0, . . . , N 1,
xi In 0 i , i 0, . . . , N 1, N ,
ui i si i , i 0, . . . , N 1
for all s N1 and all d N
December 2005
(20)
(21)
3139
v N , s
VN , s, v
(22)
x x*, s Qx x*, s u
N1
V N , s, v :
i0
u *, sT Rui u *, s xN x *, sT PxN x *, s
(23)
with the matrices Q nn , R mm and P nn
positive denite. The vectors x i n and u i m are dened
as
x 0 x I n 0
x i1 Ax i Bu i x x d ,
uk k sk *0 k, sk
(24a)
i 0, . . . , N 1
X Nv : x x, d 2n such that
N , s A for all s
(24b)
u i u * , s K x i x * , s i,
i 0, . . . , N 1
(24c)
(25)
As is standard in receding horizon control, for the current
augmented state and set point s, we only keep the rst
element *0 ( , s) of the solution to the FHOCP. Using this
receding horizon principle, we dene our receding horizon
control input as
3140
(26)
December 2005
(27)
(28)
AIChE Journal
(30)
c : b min Gd d min Gs s
(31)
where
d N
s N1
k3
c b absGd 1 absGs 1
December 2005
(32)
N1
v*, s argmin
v
iT Wi Fv c H Hs s
i0
(33)
where W : R B T PB as in Lemma 4.
Illustrative Example
Process and constraints
As an example, we consider a jacketed continuous stirred
tank reactor (CSTR) studied by Henson and Seborg40 in which
an irreversible liquid-phase reaction occurs. A detailed nonlinear model has two states (reactant concentration and reactor
temperature), one input (cooling liquid temperature) and two
disturbances (feed temperature and feed reactant concentration). This CSTR shows three steady states, two of which are
open-loop unstable, and for quality and safety reasons the
middle conversion open-loop unstable steady-state is chosen as
a nominal operating set point. Using a sampling time of t s
0.1 min and introducing deviation variables (from the corresponding steady state) a linearized model is as follows
Vol. 51, No. 12
3141
W Fv c H (0) H s(0)
N1
x(0)) dT d)
T
i
i0
v
Figure 2. Domain of attraction (XN
) for different xed
horizons.
0.7776 0.0045 x k
xx kk 11 26.6185
1.8555 x k
0.0004
0.0002 0.0893 d k
0.2907 uk 0.1390 1.2267 d k
1
in which 1,000.
For the receding horizon controller based on Q I 2 and
R 0.2, the plant state sequence, x, is also shown in Figure
2. Notice that the state sequence x initially starts at the
boundary of the domain of attraction X v10 and enters the domain
of attraction of the linear controller X 0 in nite time. As
expected from Theorem 4 the proposed controllers asymptotically drive the controlled variable to the asymptotic set point
despite the presence of persistent unmeasured disturbances.
Also, when the set point is changed the controllers drive the
controlled variable to the new set point. Moreover, it is interesting to notice that the choice of penalty matrices has a direct
impact on the closed-loop performance. As expected, when a
lower input penalty R is chosen, the disturbance is rejected
x 1k
zk 0 1 x k
2
1 s 1
December 2005
AIChE Journal
[0, 4)
[2
0.1] T
0
[4, 8)
[2
0.1] T
0
[8, 12)
[2
[12, 16)
0.1] T
1
0.1] T
[2
1
[16, 24)
[2
0.1] T
1
[20, 24]
[2
0.1] T
1
(and the set point is reached) more quickly and a larger control
input is used.
In Figure 4 we present the closed-loop simulation results
obtained with the same four controllers as in Figure 3 when the
plant is described by the original nonlinear system.40 From
these results it is clear that the proposed controllers are able to
achieve offset-free control even when there is a mismatch
between the plant model, and the actual plant dynamics. As
mentioned after Lemma 3, as long as the closed-loop system is
stable, it is possible to show that offset-free control holds
independently of the actual plant dynamics.
We nally present in Figure 5 a comparison of the proposed
receding horizon controller with a standard (that is, nonoffset-free) robust receding horizon controller. As an example we
chose the approach in,17 which is similar to the one proposed in
this article, in the sense that a prestabilizing gain matrix is used,
and the plant state prediction at the end of the horizon is
restricted to be in the maximal disturbance invariant set .
Both controllers are based on the same stabilizing gain matrix
K, which is the optimal LQR gain with Q I 2 and R 0.2.
AIChE Journal
December 2005
Conclusions
This article has shown how one can design a dynamic
receding horizon controller that guarantees robust constraint
satisfaction, robust stability and offset-free control in the pres-
3143
December 2005
lizing control law, one could consider optimizing over feedback policies.2,3,15,19,22 This will enlarge the region of attraction
of the receding horizon controller at the expense of an increase
in computational complexity.
The important problem of guaranteeing robust stability,
performance, constraint satisfaction and offset-free control
when output feedback (rather than state feedback) is used,
remains to be addressed.
Literature Cited
1. Blanchini F. Set Invariance in Control. Automatica. 1999;35:1747
1767. Survey paper.
2. Mayne DQ. Control of constrained dynamic systems. Euro J of Control. 2001;7:8799. Survey paper.
3. Mayne DQ, Rawlings JB, Rao CV, Scokaert POM. Constrained model
predictive control: Stability and optimality. Automatica. 2000;36:789
814. Survey paper.
4. Findeisen R, Imsland L, Allgower F, Foss BA. State and output
feedback nonlinear model predictive control: An overview. Euro J of
Control. 2003;9:190 206. Survey paper.
5. Morari M, Lee JH. Model predictive control: Past, present and future.
Computers & Chem Eng. 1999;23:667 682. Survey paper.
6. Blanchini F. Ultimate boundedness control for uncertain discrete-time
systems via set-induced Lyapunov functions. IEEE Trans Automatic
Control. 1994;39:428 433.
7. Mayne DQ, Schroeder WR. Robust-Optimal Control of Constrained
Linear Systems. Automatica. 1997;33:21032118.
8. Bemporad A, Casavola A, Mosca E. Nonlinear Control of Constrained
Linear Systems via Predictive Reference Management. IEEE Trans.
Automatic Control. 1997;42:340 349.
9. Bemporad A. Reference governor for constrained nonlinear systems.
IEEE Trans. Automatic Control. 1998;43:415 419.
10. Bemporad A, Mosca E. Fullling hard constraints in uncertain linear
systems by reference managing. Automatica. 1998;34:451 461.
11. Gilbert EG, Kolmanovsky I. Nonlinear tracking control in the presence
of state and control constraints: A Generalized Reference Governor.
Automatica. 2002;38:20632073.
12. Gilbert, EG, Kolmanovsky I, Tan KT. Discrete-time reference governors and the nonlinear control of systems with state and control
constraints. Int J of Robust and Nonlinear Control. 1995;5:487504.
13. Badgwell TA. Robust model predictive control of stable linear systems. Int J of Control. 1997;68:797 818.
14. Bemporad A. Reducing conservativeness in predictive control of constrained systems with disturbances. In: Proc. 37th IEEE Conference on
Decision and Control; Tampa, FL. 1998.
15. Bemporad A, Borrelli F, Morari M. Min-max control of constrained
uncertain discrete-time linear systems. IEEE Trans Automatic Control.
2003;48:1600 1606.
16. Chisci L, Zappa G. Dual mode predictive tracking of piecewise constant references for constrained linear systems. Int J Control. 2003;
76:6172.
17. Chisci L, Rossiter JA, Zappa G. Systems with persistent disturbances:
predictive control with restricted constraints. Automatica. 2001;37:
1019 1028.
18. Kerrigan EC, Maciejowski JM. On Robust Optimization and the
Optimal Control of Constrained Linear Systems with Bounded State
Disturbances. In: Proc. European Control Conference: Cambridge,
U. K.; 2003.
19. Kerrigan EC, Maciejowski JM. Feedback min-max predictive control
using a single linear program: Robust stability and the explicit solution. Int J of Robust and Nonlinear Control. 2004;14:395 413.
20. Lee YI, Kouvaritakis B. Constrained receding horizon predictive control for systems with disturbances. Int J Control. 1999;72:10271032.
21. Mayne DQ, Langson W. Robustifying Model Predictive Control of
Constrained Linear Systems. Electronics Lett. 2001;37:14221423.
22. Scokaert POM, Mayne DQ. Min-Max Feedback Model Predictive
Control for Constrained Linear Systems. IEEE Trans Automatic Control. 1998;43:1136 1142.
23. Rawlings JB, Meadows ES, Muske KR. Nonlinear Model Predictive
Control: a Tutorial and Survey. In: AD-CHEM Conference: Kyoto,
Japan; 203214 1994.
AIChE Journal
(A3)
(A5)
Proof of Theorem 2
Robust constraint satisfaction follows immediately from the
fact that is robust positively invariant for the closed-loop
system (Eq. 15), and the fact that is constraint-admissible.
Since is strictly stable, (I ) 1 exists and hence
is well-dened and unique. Note also from the proof of Lemma
3 that : limk3 (k). Robust asymptotic stability
follows from Theorem 1 by dening
: ,
w : d d s s
Proof of Lemma 3
Since limk3 s(k) s and limk3 d(k) d we have from
Eqs. 14 and 15 and from Lemma 2 that
: lim k d s u d ,
k3
(A1)
in which u s . Let be partitioned as follows:
( x , x , d ) in which each block is a column vector of
length n. We can rewrite Eq. A1 explicitly as follows
x Ax Bu Ed
(A2a)
x Ax Bu x x d
(A2b)
d x x d
(A2c)
AIChE Journal
December 2005
Proof of Theorem 3
Although a result, similar to the one stated here, appears to
be known 17, Sect. 4.2, we have been unable to nd a detailed
proof in the literature. Classical robust open-loop receding
horizon control,3 Sect. 4.5 is well-known to exhibit infeasibility
problems if the plant is open-loop unstable and no prestabilizing policy is used in the predictions (Eq. 22). However, it is a
remarkable fact that one can remove this problem by optimizing over a sequence of perturbations to a prestabilizing control
law. To show that this is indeed still true for the control
algorithm proposed in this article, we present a detailed proof.
It follows trivially from Assumption 5 that X 0 contains the
origin in its interior. The rest of the proof is by induction.
Let the plant state x X vj , where j {1, . . . , N 1}, the
controller state ( x , d ) be such that j ( , s) is nonempty and vj
: ( 0 , . . . , j1 ) j ( , s) be an admissible perturbation
sequence of length j. Also, let sj1 : (s 1 , . . . , s j1 ) j1
and dj : (d 0 , . . . , d j1 ) j be allowable set point and
disturbance sequences of length j 1 and j, respectively.
Vol. 51, No. 12
3145
sequence { (k)} k0
: {0, 0, . . .}.
It follows that if j for all sj1 j1 , and all dj
i , then j1 for all si i , and all di1 i1 .
This implies that if vj j ( , s), then (vj , 0) j1 ( ,
s). Hence, if j ( , s) is nonempty, then j1 ( , s) is nonempty. It follows from the denition of X vj that if x X vj , then
v
v
x X j1
, hence, X vj X j1
.
Using similar arguments as above, the result is completed by
noticing that X 0 X v1 .
Proof of Theorem 4
Sufciency. Suppose that x(0) X Nv , then it immediately
follows from Eq. 27 that for any initial set point s(0) one
can choose a controller state ( x (0), d (0)) such that N ( (0),
s(0)) A, and, hence, the FHOCP (Eq. 22) has a solution.
This implies from Lemma 5 we have that N ( (k)) A for all
k , and also that
: lim k : lim *0 k, sk 0
k3
3146
lim k d s d s
k3
u d
in which u s s .
The fact that Eq. 4b holds follows trivially from Lemma 5,
and the denition of N .
Necessity. This is obvious because if x(0) X Nv , then we
either have that x(0) , or that there exists an s(0)
such that for all ( x (0), d (0)) 2n , N ( (0), s(0)) A,
and, hence, the control input is undened at time 0.
Finally, robust asymptotic stability follows from Theorem 1
and can be shown in a similar fashion as in the proof of
Theorem 2. This is because it is easy to show that for any
, the optimal perturbation *0 ( , s) 0 for all s .
Hence, we can write the closed-loop system dynamics in a
neighborhood of , in terms of the shifted variables, as (k
1) (k) w(k). Again, the proof is completed by noting
that limk3 w(k) 0.
(A6)
k3
December 2005
AIChE Journal
Inventory control is based on the idea manipulating process ows so that the inventories
follow their set points. The operator mapping ows to inventories in a coarse grained
(macroscopic) system is passive and any input strictly passive (ISP) feedback controller can,
therefore, be used to achieve input-output stability. Examples of ISP controllers include the
PID controller, parameter adaptive feedforward control, optimal controllers and many
nonlinear and gain scheduling controllers. Input-output stability and convergence of inventories, such as total mass and energy can be used to show that other internal state variables
are bounded and belong to an invariant set. However, this does not necessarily imply
convergence of all state variables to stationary valves because a stabilizability condition,
known as strict state passivity, must be satised. The 2nd law of thermodynamics is used to
develop sufcient conditions for strict state passivity in the space of intensive variables. The
theory relies on following two assumptions concerning nonequilibrium systems: (1) The
hypothesis of local equilibrium, and (2) That a local entropy is dened using semiclassical
statistical mechanics. These assumptions allow us to dene a local entropy function for the
coarse grained system which is homogeneous degree one, concave, and has positive temperature. Subject to these conditions stability theory is developed for conjugate variables, such as
temperature, pressure and chemical potential in innite dimensional reaction-diffusion-convection systems. In classical irreversible thermodynamics this type of analysis assumes
linearity and symmetry of transport relations. The use of Gibbs tangent plane condition allows
us to dene a Lyapunov like storage function for passivity design, which gives stability criteria
for nonlinear problems. The resulting sufcient condition for stability can be expressed in
terms of dimensionless groups, similar to the (second) Damkohler number. Simple simulation
examples illustrate the application of the theory. 2005 American Institute of Chemical Engineers
AIChE J, 51: 3147-3166, 2005
Introduction
In the context of control of distributed parameter systems in
chemical engineering, four main categories of control problems
Correspondence concerning this article should be addressed to B. E. Ydstie at
ydstie@andrew.cmu.edu.
AIChE Journal
December 2005
3147
December 2005
recently El-Farra (2003)20 extended the method to cover systems with input constraints, whereas Kazanzis and Kravaris
(1999)21 related the energy based method to predictive control.
Sliding mode control (Palazoglu and Owens, 1987; Hanczyc
and Palazoglu, 1995; Sira-Ramirez, 1989)22,23,24 has been applied to control TRPs. The main idea behind this technique is
to let a discontinuous feedback controller dene a so-called
sliding surface on which the system has desirable behavior. A
Lyapunov-like stability condition guarantees that the distance
to the sliding surface decreases exponentially along all system
trajectories. Application of sliding mode control, a form of high
gain feedback (the feedback gain is innitely large), requires
that the input-output pairing (or transformations thereof) are
passive. Finally, Godasi et al. (2002)25 develop a symmetry
group representation which they use to develop control methods for PDE systems.
The inventory approach for stabilization of TRPs distinguishes itself from the methods mentioned previously in that
we develop stabilizing controllers using macroscopic (integrated) properties for feedback. The method is related to the
averaging methods used to study nonlinear systems and stochastic processes since the use of inventories leads to smearing of the detailed structure of the dynamics. However, in the
inventory approach we obtain the reduced order model by
averaging over space rather than time. This gives a coarse
grained view of the system dynamics and in this way the
method can be linked to the classical methods of irreversible
thermodynamics and statistical mechanics. The resulting manifold is described by a small number of ODEs (one for each
control objective) derived from the macroscopic conservation
equations. The nal control design is achieved using either
Lyapunov or passivity design techniques. A description of this
approach and its application to nite dimensional systems with
and without input constraints is given by Farschman et al.
(1998).5 Ydstie and Jiao (2004)26 applies inventory and ow
control to a oat glass production system and Ruszkowski et al.
(2005)27 applied inventory and ow control to real-time optimization of a silicon furnace.
Passivity, a concept derived from electrical circuit theory, is
encountered in many scientic disciplines (Desoer and Vidyasagar, 1975).28 The fundamental idea is based on dissipation
of useful work by irreversible transformations and input-output
pairings with phase shift no more than /2. Willems
(1972a,b)29,30 developed a systems perspective for dissipativity
and linked the concept to state space representations. Byrnes et
al. (1991)31 showed that passivity and Lyapunov stability is
equivalent for a class of feedback systems using geometric
methods. Van der Schaft (1996)32 developed control methods
linking passivity and L2 stability, while Slotine and Li (1992)16
and Ortega et al. (1998, 2001)33,34 stressed passivity based on
physical arguments following from the law of conservation of
energy. Krstic et al (1995)35 linked passivity and nonlinear
adaptive control. Ydstie and Alonso (1997)36 advanced the idea
of combining thermodynamics and passivity and showed that
passivity could be motivated using a storage function related to
the available work, and Gibbs tangent plane criteria for phase
stability. Alonso and Ydstie, (2001)37, Ydstie (2002)38, and
Coffey et al. (2000)39 highlighted the importance of concavity
of the entropy function, and used the Poincare inequality to
estimate the impact of the size on stability of the system.
The main objectives of the current article are: (1) to give a
Vol. 51, No. 12
AIChE Journal
Classical Thermodynamics
n2
Let , a convex subset of R
, be called the space of extensive variables and let Z denote an arbitrary point. We can for
example have Z (U,V,M1,. . .,Mn,X)T where U is the internal
energy, V is the volume, Mi is mass or moles of chemical
component i and X can correspond to the charge, degree of
magnetization, area, momentum et cetera.
Assumption A1 (The structure of equilibrium thermodynamics): There exists a C2 function, S : R, called the
entropy so that1:
1. For any positive constant we have S(Z) S (Z) (S
is positively homogeneous of degree one).
2. For all points Z 1 , Z 2 and any positive constant
we have
S Z 1 1 Z 2 SZ 1 1 SZ 2
(S is concave).
3. T (U/S) 0(the temperature is positive).
The developments below do not depend on a particular
denition of S since we do not actually need to evaluate the
entropy function numerically. It sufces that there exists an
entropy function which satises the axioms given earlier. We
also want to note that the assumption of concavity is not needed
for passivity design. It sufces that the entropy is over-bounded
by a concave function. This is demonstrated in the rst part of
the article by Ydstie and Alonso (1997).36
To make the discussions precise, consider the example of
semiclassical statistic mechanics. The microcanonical partition
function for a single component system of N structureless
particles is dened so that (Ruelle, 1969)40
U, V, N
3NN!
AIChE Journal
December 2005
and Dorfman, 1971).43 Methods to evaluate (U, V, N) numerically have recently been developed by Wang and Landau
(2001) and Shell et al. (2002).41,42
The implied function relating Z (U,V,N) to S given by the
two equations earlier is called a fundamental relation. In the
case of an ideal gas with constant heat capacity the result is
very clear since we get an explicit formula for the entropy so
that
S 2 S 1 c Vln
U2
V2
R ln
U1
V1
3149
dS
S Z 1 1 Z 2 SZ 1 1 SZ 2
It follows that S is concave. The reverse argument follows
immediately.
We now introduce the set of conjugate variables
w
S
Z
(1)
P
1
1
n
dU dV
dM 1, . . . , dMn
T
T
T
T
MZ 0
(4)
where
M ij
(3)
Using Eq. 3, and the chain rule we can now write (S/Z)
(w/Z) Z w T . It follows from Eq. 1 that
(2)
December 2005
2 S
0
Zi Zj
(5)
AIChE Journal
0 dw TZ d
P
1
V d
U
T
T
1, . . . , d n M
n
M
d
T
T
AIChE Journal
dv
p
dt
v is here the vector of coarse grained variables (inventories)
corresponding to the total internal energy, volume and number
of moles (mass) of each species so that
v T V t, U t, M 1t, . . . , M nt
(7)
n
and M i1
M i denotes the total mass (moles).
Neither the Helmholtz function F UTS, which is based
on xing the micro-canonical ensemble (U,V,N) nor Gibbs
function G H TS, which is based on xing the canonical
ensemble (T,P,N), provide convenient starting points for the
development of process control theory. Both functions fail for
similar reasons. The Helmholtz function fails because here is
no way to guarantee that N and U are xed unless the system
already is well controlled. The Gibbs function fails because we
need to develop methods to stabilize pressure and temperature
before we can use the stability theory. In practice these problems show up as a problem of lower-boundedness of G and F
in unconstrained systems (Coleman and Owen, 1974).46
Ydstie and Alonso (1997)36 proposed to overcome the problem of lower-boundedness by dening a Lyapunov function
which measure distance to a xed reference point in a stationary, open system. They used the microcanonical ensemble
description, and the concavity of the entropy as their point of
departure. This theory led to denition a Lyapunov function
related to thermodynamic availability and exergy dened so
that for any pair of points Z1, Z2 in we have
A 1Z 2 SZ 1 w 1TZ 2 Z 1 SZ 2 0
(6)
(8)
December 2005
f
m
i1
vt v0
pds
3
Note that A does not dene a metric on the space * since we do not necessarily
have symmetry. Consider two arbitrary points Z1 and Z2. We may then have A1(Z2)
A2(Z1), indicating that the distance measured in one direction is not equal to the
distance measured in the opposite direction.
3151
can furthermore calculate the gradients of intensive variables which are needed to estimate the rate of energy and
mass ow between the subregions. HLE is very well tested
and holds even at very small scales and quite sharp gradients.
According to the denitions above we can approximate the
coarse grained system locally (Eq. 9) by a system of partialdifferential equations on the form
f
z
t
x
t fL, t f0, t
As illustrated in Figure 4 we see that in order to predict the
evolution of a process from its initial condition we need to
evaluate a line integral (the action) involving three classical
branches of chemical engineering: thermodynamics, transport
and reaction engineering.
In the coarse-grained system we have
v it
x, t zi x, tdVt i 1, . . . , n 2
x, tdx
i1
and pt
(14)
where
fconv uz
(15)
Vt
(13)
This means that the local state z(x,t), and the specic properties Z are equivalent on small enough scales. We can
then use statistical/quantum mechanics to develop fundamental equations that link the energy and mass conservation
laws with thermodynamic quantities like temperature, pressure and chemical potentials. In course grained systems we
3152
December 2005
U
V
z ,
M1
2
M
U
V
f conv m
,
M1
2
M
Q
0
fdiff ,
FD1
FD2
u
u
2
x
x
u
x
1
2
(16)
the heat
We note that m
denotes the center of mass ow, Q
ow FDi the rate of diffusion, is the viscosity as used in the
Navier-Stokes equations and i is the local rate of chemical
reaction. The conservation law corresponding to volume (the
second entry) does not add information and is only introduced
for convenience4. The term P(u/ x) is the generation of
internal energy due to compression. This term is reversible and
4
The corresponding conservation law is given by ( V / x) (m
V / x)
(u/ x) which is true but uninteresting.
AIChE Journal
i x, t i0T, P k BT lnai zi
where ai is the activity coefcient. A continuous relationship,
therefore, exists among concentrations, pressures, temperatures
and chemical potentials. Reaction rates expressed in terms of
concentrations can, therefore, be related to the chemical potentials and more generally to the afnities as shown in the
Appendix. For example, the rst-order reaction A 3 B with
Arrhenius rate expression can be written so that
x, t k 0expE/RTx, tcA x, t
(17)
x, t
w w 1 w 2
X X 1 X 2
t
x
using the function A as a Lyapunov like storage function. This
theory can be extended using the idea of contraction analysis as
introduced by Lohmiller and Slotine (1998)51.
Lemma L1 (Evolution equation for A): Suppose that assumptions A1 and A2 are satised. We then have
A fTw
T
f diff
X wT
t
x
Proof: Since A1 and A2 are satised we can use Eq. 13
and set z( x, t) Z ( x, t). We can now write
mc S
K s c S c S2/K i
A
1 z1 2 z2
w1 w2 T
t
t
f
A
wT
t
x
w 1 w 2 k 0w 1 w 2
It follows that
w 1 w 2 Tw 1 w 2 k0 w1 w2 2
A fTw
f TX w T
t
x
w
x
(18)
Wt
f f 1 f 2
1 2
5
In the general case the right hand side of the energy balance consists of additional
n
terms P: u i1
f diff i Fi where the rst term represents the conversion of
kinetic energy to internal energy through compression and viscous dissipation, and the
second term represents the conversion potential energy to internal energy through
diffusion.
December 2005
A x, tdx 0
(21)
AIChE Journal
(19)
fT X wT dx
3153
and
Kw
wT wT Kw 0
and
X TX dx
f TX w T dx f Tw 0L 0w Tw
and 2 fT X 0 XT X
w Tw dx
2
kL 0
L2
2
dWt
f Tw 0L 0 0 2 Wt
dt
L
X TX dx 0
2
L2
December 2005
w Tw dx
0 0
2
kL
L2
2
L2
AIChE Journal
Wt W0
uyds
s 3s 0.5
s 10s 0.5s 4
(22)
December 2005
3155
1. passive if 0.
2. strictly input passive if u and 0.
3. strictly output passive if y and 0.
4. strictly state passive if represents the state and 0.
The notation 2 denotes the L2 norm for a function dened
on a domain so that for square integrable vectors we have
22
T d
w t, x Tw t, xdx
(23)
f t, 0
u
f t, L
w t, 0
y w t, L
Thus, the input vector consists of ow deviations at the boundary and the measurement consists of intensive variable deviations at the boundary. Strict state passivity (in the space of
intensive variables) follows when we make the assignment w
. We note that passivity has a very nice intuitive interpretation which is evident from inequality (Eq. 23). The input and
the output in the average have the same sign. In other words,
if u is positive then y should be positive most of the time.
Conversely, if u is negative then y should be negative most of
the time. In the nonlinear case this translates into a sector
condition which plays a crucial role in nonlinear stability
theory.
One aspect concerning passive systems is that they are easy
to control. This can be observed by dening a proportional
control
u Ky,
Wt W0
Ky 2dt 0
w Tw dt
K0
December 2005
6
Note that convergence is in the L2 sense, meaning that the error signals are square
integrable. This result can be strengthened to one of asymptotic stability of the signals
have uniformly bounded rst derivative. It turns out that it it quite easy to establish
that this is the case if the smallest eigenvalue of L is bounded away from zero.
AIChE Journal
Wt W0
e Td CeT edt
eT edt W0
eT ddt
eT edt W0
dT ddt
eT edt
Figure 8. Inventory controller acts on transformed measurements and generates a set of synthetic
control outputs which need to be transformed
to real process inputs through control variable
synthesizers (CVs).
Hence
W0
eT edt
eT edt
dT ddt
is passive.
Proof: Let
e Tedt/
d Tddt
1
C
(24)
AIChE Journal
dv*
dt
(25)
up
1
v v* Tv v*
2
e v v*
December 2005
V e Te e Tv v * e Tu
Passivity follows.
The input output pair is called synthetic since it does not
necessarily correspond to what is directly measured and manipulated in the real process.
Denition D2 (Inventory Control): A control strategy
which ensures that an inventory asymptotically tracks a desired
set point is called inventory control.
It follows from Lemma L2 and Result R4 that inventory
control is input/output stable when we use feedback-feedforward control in the form
u m, z, d p m, z, d
dv*
Ce
dt
(26)
3157
u Kc e
1
s D e
I
eu Kc e2
m, z, d p z, d 0
has a unique inverse with respect to the manipulated variable
m. Farschman et al. (1998)5 refers to this property as
controllability. This condition fails close to maximum conversion point for component B in reaction scheme A 3 B 3 C. A
further discussion is provided by Ruszkowski at al (2005).27
There are many examples of control that at are strictly input
passive.
Result R5: The PID control
m, d p m, d
3158
1
dv*
Kc e
dt
I
edt D
Kc
es Kc D ee
I
(27)
Kc
ss K c Dee
T
(28)
m, z, d, t pm, z, d, t m, z, d, t T
(29)
de
dt
December 2005
Hence
eu K ce 2
1
es D ee
I
dv*
Ce
dt
d
gm, z, d, te
dt
Vol. 51, No. 12
AIChE Journal
T te eu e T t ev *
(30)
1
t T t
2g
By differentiating V we get
and
limE U* 0
t30
V g 1 t T t
t30
M
n
V t T m, z, d, te
and E U K P
i1
All these variables are bounded from below (by zero, with
exception of U which can be dened relative to an arbitrary
reference point). It follows that there exist invariant sets Z
so that limt3 Z Z. By continuity and concavity we conclude
that similar invariant sets are dened for the intensive variables. The sets are especially simple for the total mass and
energy since they are determined by the setpoints. It is well
known that the invariant sets corresponding to the remaining
variables may be quite complex, especially for the case of
chemical reactions.
We now have the following results
Lemma L3: The process system with input and output pair
dened so that
u f0 f L
dx
dv*
dt
m, z, t pm, z, t
y M M*
dv*
Kc v v*v v*
dt
This gives the high gain for large errors and small gain for
small errors. Strict input passivity is easy to establish. Finally,
we can implement any kind of gain scheduling (hybrid) controller
f TX w T dx f Tw 0L 0w Tw
dv*
C i e,
m, z, t pm, z, t
dt
AIChE Journal
i 1, . . . , N.
December 2005
Proof:
3159
YWV
where W is dened by Eq. 21 and V by Eq. 25. We get, using
Corollary C1 and Result R4
dY
f Tw 0L
dt
Application to 1D TRPs
fTX w T dx
v v* T p
dv*
dt
w Tw dt c 0
and
eT edt c1
kc
where k is constant and c is the local concentration of chemical
specie A. In this example there is only one state variable.
Ydstie and Jiao (2004)26 and Ruszkowski et al. (2004)27 apply
the inventory control method to more complex examples.
Result R8: The transport-reaction system (Eq. 22) with
input and output dened so that
u f0 f L
e M M*
t3
limP1 P* 0
limi1 *i 0
t3
t3
t3
dx
(31)
is strictly passive if
i 1, . . . , n
limwx* wx, t 0
December 2005
D2
k 0
L2
f0 f L
dx Kc M M*
dv*
,
dt
Kc 0
(32)
Vol. 51, No. 12
AIChE Journal
The control law is quite general in the sense that we are free to
choose manipulated variables including ow rate, inlet composition or the total reaction rate. In our particular case the inlet
ux f (0) admits the decomposition
f 0 vc0 v*c*0
We can choose to manipulate the transport velocity (v) or the
inlet concentration (c(0)). Regardless of the choice, for the
control law (Eq. 32) to be solvable, specic criteria must be
satised. The requirements are as follows:
1. The righthand side of control law (Eq. 32) must be well
dened and invertible with respect to the manipulated variable
so that the control action can be implemented.
2. The geometry of the system must be chosen so that the
system is strictly dissipative. This can always be achieved by
choosing L small.
The rst point ensures that a control action can actually be
calculated; this may set some limitations on selection of the
manipulated variables as we discuss a companion article (Ruszkowski et al., 2004).27 The second point ensures that all states
(controlled and uncontrolled) reach a stable stationary point.
In the following examples we do not include the term dv*/dt
in the control laws.
Example 1. In this example, we perform a servo control
study. A step change is induced to the inventory set point, and
the transport velocity (v) is chosen as the manipulated variable.
The control law (Eq. 32) takes the form
vc0 v*c*0 vcL v*c*L
*dx Kc M M*
1
Kc (M M*)
c0 cL
dx b*
b* v*c*0 v*c*L
*dx
1
1
Kc (M M*)
c0 cL
I
(M M*)dt
dx
Figure 10. Servo control simulation results under proportional feedback and nonlinear feedforward control for Example 1.
The graph on the left shows the controlled output (Inventory
y MA). The graph on the right shows the manipulated
input (Velocity u v).
AIChE Journal
December 2005
3161
Figure 12. Servo control simulation results under proportional integral feedback and nonlinear
control for Example 1.
c0 cL
1
Kc (M M*)
v
dx b*
Figure 13. Servo control simulation results under proportional feedback and nonlinear feedforward control for Example 2.
The graph on the left shows the controlled output (Inventory
y MA). The graph on the right shows the manipulated
input (Inlet concentration u c(0)).
3162
December 2005
Figure 14. Disturbance rejection simulation results under proportional feedback and nonlinear
feedforward control for Example 3.
The graph on the top left shows the controlled output (Inventory y MA). The graph on the top right shows the
disturbance input (velocity d v). The graph on the bottom
left shows the manipulated input (inlet concentration u
c(0)). The graph on the bottom right shows the outlet concentration c(L).
Conclusions
In this article we have proposed a method to integrate
physics and process control by developing a thermodynamic
based approach to passivity based control of transport-reaction
systems. The class of systems we consider is quite broad and
encompasses a wide range of classical processes in chemical,
biochemical and metallurgical process engineering, as well as
control of small-scale systems, as long as the Boltzman principle holds and coarse graining can be applied. We provided a
review of the notion of passivity and related this to the second
law of thermodynamics. Using a local storage function con-
AIChE Journal
Acknowledgments
This research was generously supported by the National Science Foundation (CTS-9316572), Alcoa Inc. and ELKEM ASA. This article was
presented at the AIChE Annual Meeting in Austin TX, Nov. 2004.
In an interesting discussion in the poster session at the 2003 AIChE
meeting in Indianapolis, Professor Mike Doherty of UMass/UCSB pointed
out that stability conditions developed from the storage function A introduced by Alonso and Ydstie (199736), are going to be equivalent to those
using the Gibbs tangent plane condition.
Literature Cited
1. Christodes PD. Control of nonlinear distributed process systems:
Recent developments and challenges. AIChE J. 2001:514-518.
2. Buckley PS. Techniques of Process control. New York: John Wiley &
Sons, Inc; 1964.
3. Shinskey FG. Process Control Systems: Application, Design, Adjustment. New York: McGraw Hill; 1967.
4. Georgakis C. On the use of extensive variables in process dynamics
and control. Chem Eng Sci. 1986;41:1471-1484.
5. Farschman CA, Viswanath K, Ydstie BE. Process systems and inventory control. AIChE J. 1998;44:1841-1857.
6. Ray WH. Some Recent Application of distributed parameter systems
theory - A survey. Automatica. 1978;14:281-287.
7. Aling H, Benerjee S, Bangia A, Cole V, Ebert J, Emani-Naeini A,
Jensen K, Kevrekidis I, Shvartsman S. Nonlinear model reduction for
simulation and control of rapid thermal processing. Proceeding of
American Control Conference: Albuquerque, NM: 1997;2233-2238.
8. Balas MJ. Finite-dimensional control of distributed parameter systems
by Galerkin approximation of innite dimensional controllers. J of
Math Anal and Appls. 1986;114:17-36.
9. Christodes PD. Nonlinear and Robust Control of PDE Systems.
Berlin: Birkhauser; 2000.
10. Temam R. Innite-Dimensional Dynamical Systems in Mechanics and
Physics. New York: Springer-Verlag; 1988.
11. Brown HS, Jolly MS, Kevrikidis IG, Titi ES. Use of Approximate
Inertial Manifolds in Bifurcation Calculations. Continuation and Bifurcations: Numerical Techniques and Applications. (D. Roose, B. de
Dier, and A. Spence, eds.) Netherlands: Kluwer Acad. Publisher;
1990:9-23.
12. Smoller J. Shock Waves and Reaction Diffusion Equations. New York:
Springer Verlag; 1996.
13. Deane AE, Kevrikidis IG, Karniadakis GE, Orszag SA. Low dimensional models for complex geometry ows: Application to grooved
channels and circular cylinders. Phys Fuids A. 1991;3:2337-2354.
14. Christodes PD. Daoutidis P. Finite-dimensional control of parabolic
PDE systems using approximate inertial manifolds; J Math Anal Appl.
1997;216:398-420.
15. Kazantzis N, Good T. Invariant manifolds and the calculation of the
long-term asymptotic response of nonlinear processes using singular
PDEs. Comp and Chem Eng. 2003;29:999-1012.
16. Slotine JJ, Li S. Applied Nonlinear Control. Englewood Cliffs ;NJ:
Prentice Hall; 1992.
AIChE Journal
December 2005
17. Sontag ED. Mathematical Control Theory: Deterministic Finite Dimensional Systems. 2nd ed. New York: Springer; 1998.
18. Christodes PD, Daoutidis P. Robust Control of Hyperbolic PDE
Systems, Chem. Eng. Sci., 1998; 53:85-105.
19. Christodes PD, Robust control of parabolic PDE systems, Chem Eng
Sci. 1998;53:2449-2465.
20. El-Farra N, Armaou A, Christodes PD. Analysis and control of
parabolic PDE systems with input constraints. Automatica. 2003;39:
715-725.
21. Kazantzis N, Kravaris C. Energy-predictive control: A new synthesis
approach for nonlinear process control. Chem Eng Sci. 1999;54:16971709.
22. Palazoglu A, Owens SE. Robustness Analysis of a xed-bed tubular
reactor: Impact of modeling decisions. Chem Eng Commun. 1987;59:
213-227.
23. Hanczyc EM, Palazoglu A. Sliding mode control of nonlinear distributed parameter chemical processes. Ind. & Eng. Chem Res. 1995;34:
557-566.
24. Sira-Ramirez H. Distributed sliding mode control in systems described
by quasilinear partial differential equations. System and Control Letts.
1989; 13:171-181.
25. Godasi S, Karakas A, Palazoglu A. Control of nonlinear distributed
parameter processes using symmetry groups and invariance conditions. Comp. & Chem. Eng. 2002;26:1023-1036.
26. Ydstie BE, Jiao J. Passivity based inventory and ow control in at
glass manufacture, Proc. IEEE-CDC, Bahamas; December 14-17,
2004.
27. Ruszkowski M, Read M, Kaiser RW, Richardson P, Kern T, Ydstie
BE. Passivity based control and optimization of a silicon process. Proc.
IFAC World Congress; (M. Simandl and P. Horacek, eds.) July 4-8,
2005, Praha, Czech Republic, paper TU-M06-TO/I.
28. Desoer CA, Vidyasagar M. Feedback Systems: Input-Output Properties. New York: Academic Press; 1975.
29. Willems JC. Dissipative dynamical systems, Part I: General theory.
Arch Rational Mech. Anal. 1972a;45:321-351.
30. Willems JC. Dissipative dynamical systems, Part II: Linear systems
with quadratic supply rates. Arch Rational Mech Anal. 1972b;45:352393.
31. Byrnes CI, Isidori AJ, Willems JC. Passivity, feedback equivalence,
and the global gtabilization of minimum phase nonlinear systems.
IEEE Transactions on Automatic Control. 1991;36:1228-1240.
32. van der Schaft A. L2-gain and passivity techniques in nonlinear
control. London U.K: Springer-Verlag Lecture Notes in Control and
Information Science; 1996:218.
33. Ortega R, Loria A, Nicklasson PJ, Sira-Ramirez H. Passivity based
control of EulerLagrange systems, Berlin: Communications and
Control Engineering. Springer-Verlag; 1998.
34. Ortega R. van der Schaft A, Mareels I, Maschke B. Putting energy
back in control. IEEE Control Systems Magazine; 2001;4:18-33.
35. Krstic M, Kanellakopoulos I, Kokotovic P. Nonlinear and adaptive
control design, New York: John Wiley & Sons, Inc; 1995.
36. Ydstie BE, Alonso AA. Process systems and passivity via the Clausius-Planck inequality. Systems and Control Letts. 1997;30:253-264.
37. Alonso AA, Ydstie BE. Stabilization of distributed systems using
irreversible thermodynamics. Automatica. 2001;37:1739-1755.
38. Ydstie BE. Passivity based control via the second law, Comp. and
Chem. Eng. 2002;26:1037-1048.
39. Coffey DP, Farschman CA, Ydstie BE A. Distillation Stability Using
Passivity and Thermodynamics. Comps Chem Eng. 2000;24:317-322
40. Ruelle D. Statistical Mechanics Rigorous Results, Amsterdam: WA.
Benjamin Inc; 1969.
41. Wang F, Landau DP. Determining the density of states for classical
statistical models: a random walk algorithm to produce a at histogram. Phys Rev E. 2001;64:1-16.
42. Shell SM, Debenedetti PG, Panagiotopoulos AZ. Generalization of the
Wang-Landau method for off lattice simulations. Phys Rev E. 2002:66.
43. Kestin J, Dorfman JR. A course in statistical mechanics, New York:
Academic Press; 1971.
44. Callen HB. Thermodynamics and an introduction to thermostatics. 2nd
ed. New York: JohnWiley & Sons Inc; 1985.
45. Sandler SI. Chemical and engineering thermodynamics. 3rd ed. New
York: JohnWiley & Sons Inc; 1999.
46. Coleman BD, Owen DR. A mathematical foundation for thermodynamics. Arch Rational Mechanics. 1974;54,1-104.
3163
47. Hangos K, Cameron IT. Process Modeling and Model Analysis. New
York: Academic Press; 2001.
48. Kreuzer HJ. Nonequilibrium Thermodynamics and its Statistical Foundations. Oxford: Clarendon Press; 1981.
49. Astarita G. Thermodynamics: an advanced textbook for chemical
engineers. New York: Plenum Press; 1990.
50. Ottinger HC. Beyond equilibrium thermodynamics, Hoboken, NJ:
Wiley Interscience; 2005
51. Lohmiller W. JJ Slotine, On contraction analysis for nonlinear systems. Automatica. 1998;34:683-696.
52. Straughan B. The energy method, stability, and nonlinear convection.
Berlin: Springer-Verlag. Appl Math Sci Ser. 1992;91.
53. Struwe M. Variational methods. Applications to Nonlinear Partial
Differential Equations and Hamiltonian Systems. New York: Springer
Verlag; 1992.
54. Sontag ED. Mathematical Control Theory. Texts in Applied Mathematics 6. New York: Springer Verlag; 1990.
55. Khalil K. Nonlinear Systems. Englewood Cliffs, NJ: Prentice Hall
Professional Technical 1996.
56. Gilles ED. Network theory for chemical processes. Chem Eng Tech.
1998;21:121-132.
57. Hangos KM, Alonso AA, Perkins JD, Ydstie BE. Thermodynamic
approach to the structural stability of process plants. AIChE J. 1999;
45:802-816.
58. Ydstie BE. Distributed decision making in complex networks: The
adaptive enterprise. Comp Chem Eng. 2005;24.
59. Jillson KR, Ydstie BE. Process networks. Paper to be presented in
IFAC World Congress: Czech Republic; Praha; 2005.
60. Peusner L. Studies in Network Thermodynamics. Amsterdam:
Elsevier; 1986.
61. Wang J, Ydstie BE. Robust inventory control of process systems.
Boston: American Control Conference; 2004.
62. Ruskowski M. Passivity Based Control of Transport-Reaction Systems: Application to Silicon Production. Chemical Engineering; Carnegie Mellon University; 2003. PhD.
63. Kondepudi DK, Prigogine I. Modern thermodynamics: From heat
engines to dissipative structures. New York : John Wiley & Sons, Inc;
1995.
64. Onsager L, Machlup S. Fluctuations and irreversible processes. Phys
Rev. 1953;91:1301-1308.
65. Gelfand IM, Fomin SV. Calculus of variations. Mineola; NY: Dover
Publications; 1963.
66. Glansdorff P, Prigogine I. Non-Equilibrium Stability Theory. Physica.
1970;46:344-366.
67. Wei J. Irreversible thermodynamics in engineering. Ind and Eng
Chem. 1966;58:55-60.
December 2005
sdV
wT
t
z t
t
By using the transport Eq. 14 we get
s
f
wT
t
x
wT
t
x
x
Entropy Flux
(A1)
Entropy Generation
S f TX w T 0
(A2)
V
T
f conv
X u
M1
2
M
1/T
d
P/T
dx 1 /T
2 /T
By writing this expression out term by term and using the fact
we get
that u M
Xm
U
f conv
d 1
d P
d 1
1
V
M
dx T
dx T
dx T
2
M
Vol. 51, No. 12
d 2
dx T
AIChE Journal
S f diffXT X wT 0
(A3)
w T
1 u 1 u
P
P
T x T x
T 2 T1 xu
nr
i1
The rst two terms represent reversible work and they cancel.
From Eq. A3 we, therefore, get
T 2 T1 xu 0
nr
S f diffXT X
i1
1 u
S f diffXT X TR m 2
T x
1
nr
,...,
T
T
(A4)
where ij is the stochiometric coefcient of species si as it
enters in reaction j. For example if we have n 3 and r 1,
and reaction
2A B 7 C
f diff LX
(A5)
R Km
(A6)
1 u
T x
v T
n
Aj
ij
j 1, . . . , r
(A9)
i1
v k A v k v T
nr
nr
kj j
j1
kj j
ij
j1
(A7)
(A8)
i1
The rst term in expression (Eq. A4) represents entropy generation due to heat conduction and diffusion. The second term
represents entropy generation due to chemical reaction, and the
third term represents entropy generation due to viscous dissipation.
LIT assumes that the constitutive equations for diffusive
ow and chemical reaction are linear. This gives
S X TLX mT Km 2
j 1, . . . , r
ji i
where
m
v s 0,
n
December 2005
7
Any square matrix, say B, can be decomposed into a symmetric matrix S and an
antisymmetric matrix A where A has zeros on the diagonal. We write this decomposition so that B A S. We therefore have xTBx xTSx for any vector x. It
follows that inequality (A7) doesnt imply symmetry.
3165
kvv m
Using the steady-state equation A12 with A13 and A14, we get
nr
T
j j j
(A10)
S dX TL
d S
wT K
dx X
w
dx
j1
kvv
nr
T
j j j
(A11)
j1
Equation A10 now has the form given in Eq. A6. It follows that
K automatically is symmetric.
Lars Onsager (Onsager and Machlup, 1953),64 however,
argued that in LIT L must also be symmetric due to microscopic time reversibility (Kreuzer, 1984).48 This property
prompted Prigogine to propose the principle of minimum entropy production at steady state. He was able prove the generality of the principle using example problems. Later, we generalize the results due to Onsager and Prigogine, by showing
that in the linear case minimum entropy production and symmetry are equivalent.
Result A2: Consider an LIT system. The functional
L0 S (w, X, x, t)dx 0 is minimized at steady state if and
only if KL1 (L LT ) 0.
Proof: We have shown above that the rate of entropy
production is independent of the center of mass velocity since
the Gibbs-Duhem orthogonality allows us to write X T vz 0.
Without losing generality, we set u 0. At steady state we then
have
dLX
Kw
dx
(A12)
dX TL LXT
d S
dx X
dx
(A13)
S
w TK KwT
w
(A14)
Likewise
3166
December 2005
d S
S
w TKL 1 L LT
dx X
w
(A15)
S
d S
0
dx X
w
However, these are the Euler-Lagrange equations for the functional and it follows that is at an extremum. Suppose now
that KL1 (L LT ) 0. Then there exists w so that the right
hand side of Eq. A15 is not equal to zero. The E L gives the
necessary and sufcient conditions for an extremum for this
problem (Gelfand and Fomin, 1963),65 and it follows, therefore, that the functional is not minimized for general w.
That the extremum indeed is a minimum follows from the
fact that the entropy production with linear laws is convex
according to constraint (Eq. A7).
We see that the condition of Result A2 is satised if L is
symmetric. It is also satised for nonreacting systems since we
then have K 0. Minimum entropy production, therefore, is a
more general principle than symmetry.
Prigogine developed a local stability result by using the
second-order variation of the entropy as a Lyapunov function
and looked at small deviations from equilibrium steady states
(Glansdorff and Prigogine, 1970; Kreuzer, 1981; Kondepudi
and Prigogine, 1998).66,48,63 This theory was quite successful
for diffusion, heat conduction and many other phenomena that
can be modeled using linear rate expressions.
The linear theory was not successful for the analysis of
systems far from equilibrium. The use of linear reaction is
especially troublesome and LIT, therefore, has limited scope
for practical application (Wei, 1966).67 Another important
problem with LIT is that it does not address input-output
behavior. In order to make use of the formalism for control it
is necessary, therefore, to develop theory that can (1) deal with
instabilities that arise due to nonlinear phenomena, and (2)
address the input output behavior as we have done is the main
body of this article.
Manuscript received July 30, 2004, and revision received Mar. 8, 2005.
AIChE Journal
Keywords: emulsion polymerization, particle size distribution, observer, estimation, modeling, monitoring
Introduction
Controlling the particle size distribution (PSD) is an important task in emulsion polymerization processing. The PSD is an
important parameter that affects the polymer rheology, adhesion, optical properties, mechanical strength, and the latex
stability and is an important parameter in control of process
productivity because it determines the reaction rate. The particle size is also an important parameter for lm formation (see,
for example, Geurts et al.1). In addition, by producing multimodal latices, Guyot et al.2 were able to maximize the latex
solids content and to reduce its viscosity.
One of the difculties in controlling the PSD lies in the lack
of on-line instruments that are able to measure it continuously
and on-line. Quasi-elastic light scattering (QELS) is usually
used to measure the particle size for monomodal latices. Online spectroscopic techniques such as near-infrared or Raman
spectroscopies have also been used to measure the particle size
in polymerization processes.3 However, with respect to lightscattering techniques, the QELS technique is more efcient for
monomodal latices. Multimodal latices can be measured by
off-line microscopic techniques, transmission electron microsCorrespondence concerning this article should be addressed to N. Sheibat-Othman
at nida.othman@lagep.univ-lyon1.fr.
AIChE Journal
December 2005
3167
December 2005
Modeling
Population balance equations
The general population balance equation is given by the
following hyperbolic differential equation:
, t
( , t)
d
dt
generation , t
(1)
t
(2)
termination are the total rates of entry, desorption, and termination within a single particle of size between r and r r,
respectively.
The second model is the zero one model.7 This model is
adapted only for systems with the number of radicals per
particle equal to 0 or 1. Therefore, this model is generally
available for small particles, except for monomers that propagate rapidly, such as acrylates, where small particles exhibit
pseudobulk kinetics.23 This model distinguishes particles that
have a polymeric radical (n1p), particles that have no radicals
(n0), and particles that have a monomeric radical (n1m), as
shown in Figure 1 Nucleation produces particles with one
radical (n1p). A particle of type (n1p) becomes a particle of
category n0 if a new radical enters in or becomes a particle of
the category n1m if the radical it contains becomes a monomeric
radical by transfer to monomer or small chains. The distinction
Vol. 51, No. 12
AIChE Journal
between particles containing one radical and particles containing a monomeric radical is necessary in the 0 1 model because
desorption concerns only monomeric radicals.
Because the objective of this work is to monitor the PSD
starting from the beginning of the reaction, where the particles are small, the 0 1 model will be used throughout
n 0(r, t)
(r, t)[n 1p(r, t) n 1m(r, t) n 0(r, t)] k des(r)n 1m(r, t)
t
[G(r)n 1p(r, t)]
n 1p(r, t)
entr(r, t)n 0(r, t) [ r, t) k tr[M] P]n 1p(r, t) k pe[M] Pn 1m(r, t)
nuc (r r nuc)
t
r
n 1m(r, t)
entm(r, t)n 0(r, t) k tr[M] Pn 1p(r, t) n 1m(r, t)[ (r, t) k pe[M] P k des(r)]
t
(4)
nuct
Gr nuc
(5)
with
jcrit1
nucmic
em,i
[Micelle][IM i]V aq
iz
nuchom(t) k
aq
p,( jcrit1)
aq
(6)
(3)
n 1pr, t n 1mr, t
nr, t
(7)
Aqueous-phase reactions
Particle nucleation and radical entry/exit from the particles
depend on the concentration of micelles and radicals in the
aqueous phase. Therefore, to solve the population balance
equations, material balances in the aqueous phase should be
studied.
The polymerization process starts by the initiator decomposition (I) in the aqueous phase producing primary radicals (I)
that react with monomer molecules (M) to generate oligomeric
radicals (IMi). The main reactions taking place in the aqueous
phase are illustrated in Table 1.
Radical entry and desorption from particles is governed by
the coefcients (keE, ke,i, and kdes) that are assumed to be
diffusion controlled (Dw) and depend on the particle radius as
shown in Table 2. [E] is the concentration of monomeric
radicals that have a lipophilic nature and can therefore diffuse
easily into particles and desorb out of them. [T] is the total
concentration of radicals in the aqueous phase, dened by
IM E I
jcrit1
i1
Figure 1. 0 1 model.
AIChE Journal
December 2005
(8)
3169
I 2I
I M O
IM1
aq
kPI
aq
kpi
IMi M O
IMi1
Propagation
IMi T
Termination
ktaq
I T O
inactive polymer chain
ET
ktraq
kem,i
IMi Micelle O
new particle z i jcrit 1
Micellar nucleation
aq
kpj
crit 1
Homogeneous nucleation
new particle
IMj crit1 M O
IMi particlen O
particlen1
ke,i
keE
E particlen O
particlen1
1 dI V aq
aq
2k dI k pI
M aq k taqTI
V aq
dt
(9)
1 dIM iV aqt
aq
aq
k p,i1
M aqIM i1 k p,i
M aq k taqT
V aq
dt
k traqM aq k em,iMicelleIM i
1 dIM 1V aq
V aq
dt
aq
aq
k pI
M aqI k p,1
M aq k taqT k traqM aqIM 1
(10)
1 dIM iV aqt
aq
aq
k p,i1
M aqIM i1 k p,i
M aq k taqT
V aq
dt
k traqM aqIM i, i 2, z 1
1
V aq
Parameter
Description
kp MP MWm
4r2 dp NA
k des
3Dw Msat
aq
rs2 MPsat
k e,i
rs
4NaDw
k eE
k em,i
rs
3170
rmic
k e,inrdrIM i
i z, j crit 1
(12)
1 dEV aqt
1
aq
aq
V
dt
V
rnuc
(11)
1
V aq
k eErEnrdr k taqET
(13)
rnuc
4 r s NaD w
4NaDw
rnuc
z i jcrit 1
z i jcrit 1
dm
dm MPsat MWm
1/3
December 2005
AIChE Journal
dN mt
Q m k p M P
dt
(14)
RPt
4
m p(t) d pN A
3
r 3n(r)dr
rnuc
maintain the polymer particles under saturation in all the simulations ([M]P [M Psat ]). Therefore, the model representing the
concentration of monomer in the polymer particles in interval
III is not necessary.
(15)
dN St
Q St
dt
m P(t)
N m(t) N (t)
MW m
T
m
T
where Nm(t) and N m
(t) are the number of moles of residual
monomer and the total number of moles of monomer introduced to the reactor at time t, respectively; Qm is the monomer
ow rate (mol/s); mP is the polymer mass; and NA is Avogadros number. Both ways require the integration of the PSD
because , the number of moles of radicals in the polymer
particles, depends on the average number of radicals in the
polymer particles of any size at time t [n (t)] and the total
number of particles in the reactor (N PT ):
n t N TPt
NA
(16)
S aqV aq N S N SP N Sd
NSaq
N SP
nr, t N Adr
(17)
rnuc
n t
S parb sS aq
a spN A1 b sS aq
(21)
S par 4 N A
nrr 2dr
(22)
rnuc
n r, tnr, tdr
rnuc
(18)
Micelle max 0,
During interval II (saturation of polymer particles), the concentration of monomer in the polymer particles {[M]Psat (mol/
dm3)} is calculated using the partitioning coefcients18:
sat
M aq
nr, tdr
rnuc
dm
K daqMW m
where [M]sat
aq is the concentration of monomer in the aqueous
phase under saturation (mol/dm3):
sat
M sat
P M aq K paq
[Saq ] CMC
nagg
(23)
Mathematical Treatment
(20)
and n (t) can be related to n (r, t) using Eq. 7, which gives the
following:
3V d
a sdr dN A
N Sd
where N PT can be obtained by integrating the function representing the PSD [n(r, t)] over r:
N TPt
(19)
December 2005
Experimental measurement of the PSD, by CHDF for example, gives the total number of particles of size r [n(r)] and does
not give n1p(r), n1m(r), and n0(r) individually. Therefore, constructing an estimator using this measurement would be easier
Vol. 51, No. 12
3171
using a model that gives n(r) directly rather than using the 0 1
model that describes the three states individually. Therefore,
developing a model for small particles that have the same form
as the pseudobulk model would be of interest.
For this reason, the 0 1 model presented above is used in
this section to develop a new model that has the form of the
pseudobulk model.
(26)
By combining Eqs. 24 and 25 with the approximation given
by Eq. 26, the following model is obtained for n(r, t):
Grn r, tnr, t
nr, t
nuc r rnuc
t
r
Grn1p r, t n1m r, t
nr, t
nuc r rnuc
t
t
(24)
Grn1p r, t
nr, t
nuc r rnuc
t
r
(28)
r
n(r, t)
nr, t
(29)
(25)
Grn 1mr, t
0
r
This is explained by the fact that the lapse of time during which
a particle is of type n1m is very small. Actually a particle of type
n1m becomes very quickly either an n1p (if the monomeric
radical propagates), in which case it grows normally, or becomes an n0 (if the monomeric radical desorbs or terminates),
in which case it does not grow. For this reason, assuming that
the particles (n1m) grow as the particles of type n1p does not
affect the evolution of the PSD. In this case, Eq. 25 can be
written under the following form:
Using the denitions of N(r, t), n(r, t), and Eq. 24, the
expression of n1m(r, t) can be rewritten under the following
form:
n 1mr, t
entmr, tnr, t1 n r, t
r, t M Pk pe k tr k desr
k trM Pn r, tnr, t
r, t M Pk pe k tr k desr
(30)
(a)
n(r, t)
[G(r)n (r, t)n(r, t)]
nuc (r rnuc )
t
r
(b)
n (r, t)
[G(r)n (r, t)n(r, t)] n (r, t) 1
(r, t)[1 2n (r, t)]
t
r
n(r, t)
kdes r
nr, t
Nr, t
Nr, t
t
t
2
n r, t
n r, t
r, t1 2n r, t
t
3172
nr, t
(27)
(31)
ktr MP n r, t
entm r, tn r, t 1
December 2005
AIChE Journal
Numerical solution
nuct
Gr nuc
and
n rnuc , t 1
AIChE Journal
December 2005
3173
Estimation of
Excitement of equation system 31 is attributed to particle
formation either by micellar or homogeneous nucleations
[nuc(t)]. The rate of nucleation is directly related to the
concentrations of monomer ([M]Psat ), radicals ([IM i ] for z i
jcrit 1) and micelles ([Micelle]) in the aqueous phase. However, the concentrations of radicals in the aqueous phase is
sensitive to impurities and inhibitors and the concentration of
micelles depends on a number of coefcients that are not well
known (adsorption and desorption). This makes the process
irreproducible and necessitates on-line measurement of the
reaction rate and the kinetics in the aqueous phase.
In this work, the monomer conversion
Xt 1
Observer Design
A continuous stream information about the evolution of the
PSD is necessary for process control. However, the PSD is
difcult to measure on-line and the process model involves a
number of parameters that are sensitive to impurities and are
difcult to identify. This necessitates the construction of an
observer to monitor the PSD.
An observer combines the process model describing the
chemical reactions and physical behavior with the existing
on-line measurements to estimate the unknown parameters or
states of the system on-line. Among these algorithms, the
Kalman lter is often proposed in the literature. This lter
requires solving the dynamic Riccati equation, which is nonlinear. For a system of dimension k, the dimension of the space
of the Riccati equations is (k2 k)/2. Moreover, the choice of
the weighting matrices of the Riccati equations is difcult.
Therefore, the Kalaman lter becomes difcult to adjust and is
time consuming.
Our approach consists in designing an observer that derives
from the high-gain techniques (see, for instance, Bornard and
Hammouri24; Deza et al.25; Gauthier et al.26; Gauthier and
Kupka27; Farza et al.28; Hammouri et al.29; and Edouard et
al.30). The gain of the proposed observer does not require
solving any differential equation and its tuning is easier.
In the following, two estimators are constructed. The rst
one allows the estimation of the number of moles of radicals in
the particles () from the measured overall conversion. Then
this estimate is used in a second observer to estimate the PSD
3174
December 2005
N mt
N m0
is assumed to be measured on-line. This can be done by calorimetry,31-33 on-line gas chromatography, or by spectroscopy.
Based on this measurement, Othman et al.34 developed an
estimator of the parameter . The estimation of allows
detection of inhibition and determination of the real value of
the reaction rate. The estimation is based on the material
balance of monomer given by Eq. 14. In this nonlinear equation, the sole unknown parameter is . This is similar to saying,
is observable from the model 14 if the residual number of
moles of monomer is measured. A high-gain nonlinear estimator could therefore be constructed to monitor :
dN m(t)
Q m k p (t)[M] P 2 (N m N m)
dt
2
d (t)
(N N m)
dt
k p[M] P m
(32)
AIChE Journal
December 2005
3175
December 2005
AIChE Journal
n(r nuc k r, t)
x 1(t)
n(r nuc r, t)
x k(t)
xt
x k1(t)
N pT(t)
x k2(t)
nuc(t)
x k3(t)
(t)
dt
dt
rnuc
nuct Gr n r , tnr , t
(33)
ak ak1
0
0
ak1 ak2
0
0
a1
Ar 0
0
0
0
0
1
0
1
0
0
0
0
0
0
1
0
Gr n r , tnr , t 0
(34)
(t) N A
N TP(t)
n (t)
(35)
AIChE Journal
December 2005
a i Gr nuc i rn r nuc i r, t/ r
(36)
(37)
where
3
53
5
3177
The observer developed in this section thus allows us to estimate the PSD from the measurement of the monomer conversion
even under disturbances if the model of n (r, t) is supposed to be
accurate. However, the model of n (r, t) contains uncertainties
arising from inhibition and possible errors of parameter estimation. For this reason, the observer must be adapted in a way that
allows the estimation of the PSD even under uncertainties in the
model of n (r, t). Figures 15 and 16 show the inuence of n (r, t) on
the PSD. Actually, n (r, t) and therefore n (t) have a direct inuence
December 2005
AIChE Journal
Estimate (1)
n r
0.2
0.2
December 2005
3179
Figure 15 shows that n (r, t) substantially affects the nucleation rate and therefore the PSD. Figures 16a, 16b, and 16c
represent the real PSD and the PSD obtained by the constant
values of n (r, t), estimate (1) and estimate (2), respectively. This shows that in-depth knowledge of n (r, t) is necessary to obtain the real PSD. The open-loop simulation of the
model of n (r, t) that might undergo uncertainties is not ef-
Estimate (2)
n r
and thus n 1.
December 2005
AIChE Journal
k des(r)
(r, t) [M] P(k pe k tr) k des(r) (r, t) [M] P(k pe k tr) k des(r)
(38)
Br, t
December 2005
3181
Figure 17. Inuence of the parameter Dw on the parameters kdes, keE, and ke,3.
terms showed that the terms A(r, t) and B(r, t) have a slow
dynamic and do not change abruptly even under nucleation.
The second term is the most sensitive to nucleation and undergoes a rapid dynamic. This allows us to neglect changes
coming from the terms A(r, t) and B(r, t) if they are well
December 2005
AIChE Journal
t
r
n(r, t)
(39)
December 2005
3183
Conclusions
In this work, a new bulk-like model was developed to
represent the PSD in the 0 1 system. The new model is easier
to handle and makes possible the continuity between the 0 1
system (small particles) and the pseudobulk system (large
particles).
Based on this new model, a continuous-discrete observer
was developed to estimate the PSD and the average number of
radicals per particle simultaneously.
The observer is able to undergo modeling errors. An error
was supposed in the rate of desorption and absorption of
radicals, although the observer was able to detect nucleation
and to reconstruct the PSD.
This strategy should be a good tool for the control of the
PSD. Actually, the observer can be included in a control
scheme to manipulate the control variables: surfactant or initiator ow rates to obtain a target PSD.
Notation
asd surface area of droplets covered by a single surfactant molecule (4.2 1017), dm2
asp surface area of particles covered by a single surfactant molecule (4.2 1017), dm2
bs Langmuir adsorption isotherm parameter (related to the surfactant coverage area ratio) (2.1 103), dm3/mol
CMC critical micellar concentration (3 103), mol/dm3
3184
December 2005
AIChE Journal
rnuc
rd
rs
RP
nuc
nuc hom
nucmic
[Saq]
Spar
[T]
Vaq
Vd
z
(r, t)
entr(r, t)
entm(r, t)
Literature Cited
14.
15.
16.
17.
18.
19.
AIChE Journal
13.
December 2005
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
and semi-batch emulsion copolymerization of styrene and butyl acrylate. Chem Eng Sci. 2003;58:4493-4514.
Ramkrishna D. Population Balances: Theory and Applications to
Particulate Systems in Engineering. San Diego, CA: Academic Press;
2000.
Immanuel ChD, Doyle FJ III. Computationally efcient solution of
population balance models incorporating nucleation growth and coagulation: Application to emulsion polymerization. Chem Eng Sci. 2003;
58:3681-3698.
Semino D, Ray WH. Control of systems described by population
balance equationsI. Controllability analysis. Chem Eng Sci. 1995;
50:1805-1824.
Semino D, Ray WH. Control of systems described by population
balance equationsII. Emulsion polymerization with constrained control action. Chem Eng Sci. 1995;50:1825-1839.
Ohmura N, Kataoka K, Watanabe Sh, Okubo M. Controlling particle
size by self-sustained oscillations in continuous emulsion polymerization of vinyl acetate. Chem Eng Sci. 1998;53:2135.
Crowley T, Meadows ES, Kostoulas E, Doyle FJ III. Control of
particle size distribution described by a population balance model of
semibatch emulsion polymerization. J Process Control. 2000;10:419432.
Zeaiter J, Romagnoli JA, Barton GW, Gomes VG, Hawkett BS,
Gilbert RG. Operation of semi-batch emulsion polymerization reactors: Modeling, validation and effect of operating conditions. Chem
Eng Sci. 2002;57:2955-2969.
Immanuel ChD, Doyle FJ III. Open-loop control of particle size
distribution in semi-batch emulsion copolymerization using a genetic
algorithm. Chem Eng Sci. 2002;57:4415-4427.
Immanuel ChD, Doyle FJ III. Hierarchical multiobjective strategy for
particle-size distribution control. AIChE J. 2003;49:2383-2399.
Doyle FJ III, Harrison ChA, Crowley TJ. Hybrid model-based approach to batch-to-batch control of particle size distribution in emulsion polymerization. Comput Chem Eng. 2003;27:1153-1163.
Coen EM, Peach S, Morrison BR, Gilbert RG. First-principles calculation of particle formation in emulsion polymerization: Pseudo-bulk
systems. Polymer. 2004;45:3595-3608.
Bornard G, Hammouri H. A high gain observer for a class of uniformly
observable systems. Proc of the 30th IEEE Conference on Decision
and Control, Brighton, UK; 1991:1494.
Deza F, Busvelle E, Gauthier JP. High gain estimation for nonlinear
systems. Syst Control Lett. 1992;18:295.
Gauthier JP, Hammouri H, Othman S. A simple observer for nonlinear
systems: Application to bioreactors. IEEE Trans Autom Control. 1992;
37:875.
Gauthier JP, Kupka IAK. Observability and observers for nonlinear
systems. SIAM J Control Optim. 1994;32:975.
Hammouri M, Farza H, Busavon K. A simple observer for a class of
nonlinear systems. Appl Mater Lett. 1998;11:27.
Hammouri H, Targui B, Armanet F. High gain observer based on a
triangular structure. Int J Robust Nonlinear Control. 2002;12:497.
Edouard D, Schweich D, Hammouri H. Observer design for reverse
ow reactor. AIChE J. 2004;50:2155-2166.
Arzamendi G, Asua J. Copolymer composition control of emulsion
copolymer in reactors with limited capacity for heat removal. Ind Eng
Chem. 1991;30:1342-1350.
Manuscript received Jan. 5, 2005, and revision received Mar. 22, 2005.
3185
Process intensication has been the focus of increasing attention in recent years. A
volume reduction of distillation and absorption units by a factor of 100 or more using
rotating beds has been claimed. Industrial trickle-bed reactors are bulky because the
liquid ow and interphase transport of the reacting species are governed by the earths
gravity, as in the case of distillation columns. To explore the possible process intensication in these reactors, we have measured the reaction rates of hydrogenation of
-methyl styrene with palladium as catalyst in the rotating beds of spherical particles and
metal foam, which acted as catalyst support. To quantify the intensication achieved, the
reaction rates were compared with those of conventional trickle beds. The enhancement
in the reaction rates was in the range of 30 40 times in a centrifugal force eld of about
450 times the gravitational force eld. An industrial reactor of 60 m3 could be replaced
with a rotating bed 1.5 m3 in volume. It appears that there is a possibility of further
process intensication. However, the volume reduction is possible only for mass-transfer
limited reactions. 2005 American Institute of Chemical Engineers AIChE J, 51: 3186 3192,
2005
Introduction
Process intensication in distillation and absorption equipment has been the focus of increasing attention in recent years.
In conventional columns, the achievable vapor and liquid
throughputs and interphase mass transport are inherently governed by the gravitational eld. These can be enhanced in a
rotating annular packed bed [also known as Higee (high gravity)] because the liquid ows under the inuence of a centrifugal force eld of 350 1000 times the gravitational force eld.
It has been claimed that such process intensication can lead to
a volume reduction of the equipment by a factor of 100 to
1000. This in turn leads to lower capital costs and inherently
safer design because the liquid holdup is smaller than that in
conventional equipment. Ramshaw,1 at Imperial Chemical InCorrespondence concerning this article should be addressed to D. P. Rao at
dprao@iitk.ac.in.
3186
December 2005
AIChE Journal
dium. To quantify the process intensication, we have compared the rates in rotating trickle-bed reactors with those in
conventional trickle-bed reactors reported in the literature.
Figure 1. Model trickle-bed reactor.
1: Manometer tap; 2: gas outlet; 3: liquid inlet; 4: motor; 5:
gas inlet; 6: glass wall; 7: annular packed bed; 8: alumina
catalyst beads; 9: stand; 10: liquid outlet; 11: distributor; 12:
AC inverter drive; 13: cylindrical side plate with rectangular
slots; 14: cup for collecting splash liquid; 15: elliptical dish;
16: splashed liquid outlet; 17: cylindrical bafe.
December 2005
Experimental
Rotating trickle bed
Figure 1 shows the experimental rotating trickle-bed reactor.
It consisted of an outer casing, housing for the bed, a liquid
distributor, and a cup for collecting the liquid that was splashed
at the inner periphery of the bed housing. The side of the casing
was a cylindrical glass tube sandwiched between two metal
plates. To facilitate the liquid withdrawal, a stainless steel
elliptical dish was attached to the lower end of the glass wall as
shown in the gure. The housing was attached to the shaft of
the motor mounted on the top plate. The rotational speed of this
induction motor was varied continuously between 500 and
2500 rpm using an inverter drive.
The housing of the bed was made of two annular disks
separated by two inner and outer slotted cylindrical shells and
were bolted to provide a rigid housing for the bed. The dimensions of the housing were: height, 4.5 cm; inner diameter, 9 cm;
and outer diameter, 21 cm. The free area for the ow of gas and
liquid on the inner and outer plates was 50% of the total.
To ensure uniform ow of liquid over the active catalyst, the
bed of catalyst particles was sandwiched between the layers of
inert particles as shown in Figure 2a. Similarly, the metal-foam
rings were sandwiched between the layers of inert alumina
particles as shown in Figure 2b.
3187
2.2, 1.5
797.6 kg/m3
38.18%
62.80%
1.17 g/cm3
0.5366 cm3/g
1690, 2489 m2/m3
Catalyst preparation
The alumina catalyst particles were prepared by impregnating them with palladium, as suggested by Herskowitz et al.10
The visual observations indicated that the palladium was uniformly distributed throughout the particles. We have used
particles of 2.2 and 1.5 mm in diameter. The physical characteristics of particles supplied by the vendor (Norton Co., Philadelphia, PA) are given in Table 1.
The metal-foam rings were supplied by Recemat International B.V. (The Netherlands). Their structure and characteristics as supplied by the vendor are given in Table 2. The
metal-foam bed was made by stacking ve annular rings: four
of 1 cm height and one of 0.5 cm height. To deposit palladium
on these rings, we followed the electroless-deposition method
suggested by Lowenheim11 and Henry.12 First, the metal-foam
rings were given an alkaline soak for 30 min in a boiling
solution with a composition of 15 g/L NaOH and 15 g/L
Na2CO3. After soaking, the rings were thoroughly rinsed with
distilled water. These rings were then immersed in a plating
bath containing a solution of 2 g/L PdCl2, 160 mL/L NH4OH
(27%), 26 g/L NH4Cl, and 10 g/L NaH2PO4H2O at 50C. The
solution was stirred with a magnetic stirrer to eliminate gas
bubbles and to maintain uniform composition in the holes and
crevices. Upon completion of the electroless deposition, the
solution, which was dark in color, became clear (colorless).
After the deposition, the foam rings were rinsed thoroughly
with distilled water and dried in a stream of hot air. The
difference in the weight of metal rings before and after the
plating gave the exact amount of palladium deposited. The
visual observation of the metal foam indicated that it was black
in color and the palladium was uniformly deposited. The catalyst particles were activated by heating to 370C in a stream
of hydrogen for 24 h. However, the metal foams were used as
such after washing with distilled water.
Setup
The experimental setup is shown in Figure 4. Keeping the
bed rotating, rst nitrogen and then hydrogen were passed
through the casing for about 1 h each to ensure the casing was
lled with hydrogen before the start of the run. Then, the liquid
was circulated through a water bath and the bed using a
peristaltic pump (ColeParmer Instrument Co., Vernon Hills,
IL). A surge tank was used to dampen the uctuations in the
ow. The rotational speed and the liquid ow issuing out of the
RECEMAT (2733)
RECEMAT (1723)
0.6
2500
Nickelchromium
0.9
1700
Nickelchromium
102
149
0.60
0.9
102
149
0.65
0.9
Metal foams
3188
December 2005
AIChE Journal
V dC
V b dt
December 2005
3189
December 2005
AIChE Journal
that of the conventional trickle bed reactors. Because the reaction rate would be greater with larger beds and higher rpm, the
volume reduction of similar magnitude or more could be
achieved. The intrinsic kinetic rate for this reaction is 596
107 mol/(cm3 s1) at 50C with 1.0 mass % Pd and 1 atm
hydrogen pressure. This rate is about ve times the highest rate
observed. At these conditions the intrinsic kinetic regime is
being approached, and thus further intensication may not be
possible.
In the present study, the reaction rate was controlled by the
liquid-side mass transfer. For the case in which the reaction
rate is controlled by the gas-side transport, the novel rotor
design proposed by Rao et al.5 can be used to achieve process
intensication. If the rate is controlled by the reaction kinetics,
it is unlikely that the process intensication can be achieved
with a rotating trickle-bed reactor.
Conclusions
Figure 8. Comparison of reaction rates for different substrates.
, RECEMAT (1733): 1.5 mass % Pd, 360 g; , 1.5-mm
alumina beads, 1.0 mass % Pd, 485 g; F, conventional trickle
bed: 1.0 mass % Pd; - - -, Herskowitz et al.10
December 2005
Acknowledgments
We gratefully acknowledge Recemat International B.V., The Netherlands (http://www.recemat.com) for donating the metal-foam rings; Norton
Company, Philadelphia, PA for donating the alumina beads; and the partial
nancial support from Department of Science and Technology, Government of India.
Literature Cited
1. Ramshaw C. Higee distillationAn example of process intensication. Chem Engineer. 1983;389(Feb.):13-14.
2. Trent D, Tirtowidjojo D, Quarderer G. Reactive stripping in a rotating
packed bed for the production of hypochlorous acid. Process Intensication for the Chemical IndustrySmaller, Cheaper and Safer Production (Proceedings of the 3rd International Conference, Antwerp,
Belgium). Suffolk, UK: BHR Group Limited/PEP; 1999:217.
3. Rao DP, Bhowal A, Goswami PS. Process intensication in rotating
packed bed (HIGEE): An appraisal. Ind Eng Chem Res. 2004;43:11501162.
4. Sivalingam G, Radhika M, Rao DP, Rao MS. Process intensication in
a model trickle-bed reactor. Ind Eng Chem Res. 2002;41:3139-3144.
5. Rao DP, Goswami P, Bhowal A. A novel rotor design for intensication of gas-phase mass transfer in a Higee. Proc of AIChE Spring
Meeting, New Orleans, LA, April 2529; 2004.
6. Gianetto A, Specchia V. Trickle-bed reactors: State of art and perspectives. Chem Eng Sci. 1992;47:3197-3213.
7. Van Hasselt BW, Lebens PJM, Calis HPA, Kapteijn F, Sie ST,
Moulijn JA, van den Bleek CM. A numerical comparison of alternative
three-phase reactors with a conventional trickle-bed reactor. The advantages of counter-current ow for hydrodesulfurization. Chem Eng
Sci. 1999;54:4791-4799.
8. Smith JM. Chemical Engineering Kinetics. 3rd Edition. New York,
NY: McGraw-Hill; 1981:536.
9. Singh SP, Wilson JH, Counce RH, Villier-Fisher JF, Jennings HL,
Lucero AJ, Reed GD, Ashworth RA, Elliot MG. Removal of volatile
3191
10.
11.
12.
13.
14.
15.
3192
December 2005
16.
17.
18.
19.
Manuscript received Jul. 2, 2004, and revision received Mar. 29, 2005.
AIChE Journal
Introduction
Microscale devices have recently been explored for numerous applications. Examples include in situ production of toxic
and specialty chemicals, mixing, high throughput screening of
catalysts, chemical and biological agents, energy production,1-5
in situ microfabrication of structures of sizes 5 m within
small capillaries,6 chemical and biological analysis and synthesis,7-10 DNA separation and amplication using the polymerase
chain reaction,11 and even medical diagnostics. Because of
their small size, the ow in these devices is laminar and mixing
is driven only by molecular diffusion. Their small size also
translates into low surface areas for catalytic reactions.
To overcome some of the challenges stemming from device
miniaturization, ideas reminiscent of microstructured catalysts12-14 are used to fabricate microscale features in the form of
cylindrical posts (or other shapes) within a reactor to create
large catalytic surface area and possibly enhance mixing. Such
microscale posts have recently been implemented experimenCorrespondence concerning this article should be addressed to D. G. Vlachos at
vlachos@che.udel.edu.
AIChE Journal
December 2005
3193
dU
(1)
fd
U
(2)
December 2005
reported that the shape of the bluff body affects the critical Re
number. For a given length perpendicular to the ow, the
critical Re number increases with an increase of the size
parallel to the ow (such as Rec 27.8 for a at plate and
45.5 for a cylinder).
Residence times in microchemical devices range from milliseconds to a few seconds. Larger residence times are common
for slow chemistries and low temperatures, such as thermal
decomposition and reforming reactions, whereas millisecond
residence times are characteristic of fast chemistries, such as
combustion and catalytic partial oxidation reactions. Typical
post sizes in structured microdevices are of the order of hundreds of microns. For a device of length of L 1 cm and a
residence time of 1 ms, the Re number, based on a post
diameter of 300 m, (considering uid properties of air) is 75.
These back of the envelope calculations clearly show that
Vol. 51, No. 12
AIChE Journal
ow-driven oscillatory instabilities at the microscale are entirely feasible and could play an important role in enhancing
mixing, at least for the fastest reactions. When liquids are
considered, the transition to oscillatory ow occurs for slower
ows.
A limited number of experimental and theoretical studies are
available for multiple post congurations. For the case of two
cylinders, the experimental studies40-43 have focused on the
high Re number (103104) regime, whereas numerical studies44,45 have focused on the low Re number (102103) regime.
Different ow patterns have been reported depending on the
conguration. They are briey summarized here and the reader
is referred to the above references for details. A conguration
of two posts with an angle 0 or 180 is termed as inline or
tandem conguration (see Figure 1b). For a separation 3d,
no distinct vortices are seen behind the rst post. For larger
separations, vortices are seen behind both posts. The shedding
frequency is initially lower and then reaches the single post
value for larger separations. In a side-by-side conguration
(angle of 90 between the two posts; see Figure 1c), the two
posts act like a single body for spacing 1.4d. For spacing
larger than this but 2d, a bistable vortex shedding, which is
unique to this conguration, occurs with two different frequencies observed behind each of the posts. This biased ow (the
wake behind one post is larger than that behind the other)
changes intermittently. For all other orientations of the posts
(staggered conguration; see Figure 1d), the wake behind
the upstream post is found to be narrower and sheds with a
higher frequency than that of a single post. Similar but more
complex ow patterns have been observed experimentally for
three and four posts.46-49 A simple superposition principle does
not provide an adequate explanation of these ow patterns. A
complete understanding of the highly nonlinear multiple post
vortex shedding still remains elusive.
December 2005
the next larger domain (an increase by 5d along the width and
by 10d in the length) or a ner mesh (resulting from doubling
the number of nodes) was found to be 1%. The optimization
of computational effort is especially important because timedependent calculations, involving thousands of time steps, are
performed.
In most simulations, no symmetry is assumed. However, for
a limited number of cases the effect of symmetry has been
explored. Specically, for simulations with a single post where
symmetry was imposed, half of the above domain was simulated.
For multiple-post simulations, the width of the domain is
kept at 20d and the length is increased to 37.5d (this translates
to a domain of 4 7.5 mm, which is representative of a
microdevice). The nonuniform mesh discretization in this case
is 120 nodes (width) 225 nodes (length) 60 nodes (on the
surface of each post), which results in a mesh consisting of
nearly 200,000 nodes.
The commercial CFD software, Fluent (version 6.1.22)54 is
used to obtain time-dependent solutions using a nite-volume
method. The time-dependent 2-D continuity, momentum, energy, and species conservation equations are solved using a
second-order, implicit time-stepping method. A rst-order upwind discretization scheme is used for density, momentum,
species, and energy equations and a Standard pressure
interpolation scheme is used along with the Simple algorithm for the pressurevelocity coupling.54 A segregated solver
with an underrelaxation scheme is used to obtain the solution.
The equations are integrated with a time step of 0.5 s,
typically, for a total of 5000 10,000 time steps and up to
65,000 time steps. Time-dependent statistics are obtained after
3000 initial time steps to eliminate initial transients. A single
simulation, as described above, takes about 1 week of CPU
time on a single Pentium Xeon (2.4 GHz with 2GB of RAM)
processor.
The inlet gas (A), having properties of air (that is, molecular
weight of 29 and a diffusivity of 2.9 105 m2/s) is assumed
to ow in at a uniform velocity and a xed inlet temperature of
923 K. These conditions come from our previous work on
ammonia decomposition in post microreactors that ensure complete decomposition of ammonia.55 Re 100 is a standard
benchmark problem in the uid dynamics community,56 and is
also chosen here, unless otherwise stated. These reactive ow
Vol. 51, No. 12
3195
(3)
kmol m2 s1 148CA
(4)
(5)
(6)
December 2005
AIChE Journal
Figure 4. Effect of a catalytic post on oscillatory behavior for various types of surface reactions.
Trajectories of (a) the transverse velocity component (m/s)
and (b) mass fraction of species A, for the catalytic surface
reaction A 3 nB for various values of n and reaction orders
indicated. (c) Snapshot of the density contours (kg/m3) for a
rst-order reaction A 3 2B. [Color gure can be viewed in
the online issue, which is available at www.interscience.
wiley.com.]
December 2005
3197
absence of a post, the reactants mix only by molecular diffusion and the reaction is conned to a very small zone. In the
presence of a post, on the other hand, the ow-driven oscillations create an enlarged reaction zone. Figure 6a shows the
cup-averaged mass fraction of product C at the outlet as a
function of time. Signicantly higher conversions are found in
the oscillatory case, which is explained by the larger interfacial
area created by the wakes behind the post.
In summary, it is found that ow-driven instabilities significantly affect the observed reactivity. The absence of the vortex
shedding (symmetry enforced) would lead to the formation of
a single wake and higher conversions for non productinhibited
positive order surface reactions. For volumetric reactions between non-premixed reactants, the wakes enhance the contact
area and can result in much higher conversions.
Arrays of Posts
mix 1 2
elongated wake behind the post) without oscillations. This
nding clearly indicates that symmetry breaking is a necessary
condition for the formation of oscillatory wakes.
Figure 5c shows the effect of symmetry on the time-averaged conversion calculated at the reference point for various
values of n and reaction orders of the surface reaction (Eq. 3).
In all cases the time-averaged conversion in the oscillatory ow
is lower than the corresponding steady-state conversion for the
symmetric case. For the latter case, the single wake behind the
post creates a stationary region of lower velocity and higher
residence time and thus of higher conversion. The time-averaged exit conversion is also found to be higher for the symmetric case but the difference is marginal (in the decimal
places) because of the low surface area of a single post.
The case of volumetric reactions of non-premixed reactants
is studied next. Equal ow rates of non-premixed components
A and B enter the domain (as shown in the schematic in Figure
6a) and react according to the volumetric reaction (Eq. 5).
Figures 6b and 6c depict the mass fraction contours of product
C in the absence and presence of a post, respectively. In the
3198
December 2005
y i 0.5VdS
VdS
(7)
AIChE Journal
Figure 7. Contours of mass fraction of species A indicating the effect of different number of post rows and spacing
between posts on mixing.
[Color gure can be viewed in the online issue, which is available at www.interscience.wiley.com.]
Figure 8. Effect of the number of post rows and row spacing on mixing.
Panels (a)(f) depict the average mixing parameter at the exit for conditions corresponding to the contours in Figure 7. [Color gure can be
viewed in the online issue, which is available at www.interscience.wiley.com.]
AIChE Journal
December 2005
3199
is slightly reduced when more than one row of posts is used for
a relatively small spacing between rows. For this spacing,
wakes do not have room between successive rows of posts
to develop. Apparently, this post spacing does not lead to
constructive interference. As indicated in Figures 7c and 7d,
the small separation between the rows of posts (of 3d)
renders the multiple rows effectively acting as a single row.
The pressure drop in the micromixer is another important
performance criterion. Because of the relatively open conguration, a small pressure drop (2 Pa) is observed for a single
row of posts. For ve rows of posts, the pressure drop is still
small (250 Pa), indicating no additional pumping devices
will be required to overcome the loss of pressure head resulting
from the enhanced mixing achieved in these post micromixers.
Figures 7e and 7f show the species contours for increased
separation of values of 4.5d and 10d, respectively, between
two rows of posts and Figures 8e and 8f show the corresponding mixing parameter. With an increase in row interspacing
( 4.5d), a wake develops behind the rst row that causes a
constructive interference with the wake from the second row of
posts, resulting in an increase in the wake amplitude. This
further results in enhanced mixing, as seen from the mean
values in Figure 8e. Because of the larger amplitude of the
oscillations, the standard deviation, however, increases from a
value of 0.005 to 0.0184 with this increase in row spacing.
With a further increase in spacing ( 10d), the wakes behind
the rst row of posts reach the second row with a very large
amplitude, causing a destructive interference and resulting in
worse performance but still better than that of the 3d
separation case. Given the variation in the mixing performance
as a function of row spacing, the separation between the two
rows is systematically varied in integer multiples of post diameter (at Re 60 and 100). The results shown in Figure 10
indicate that mixing is a highly nonlinear function of row
3200
December 2005
AIChE Journal
December 2005
UL
D
(8)
3201
Strook et al.,71 and Eneld et al.,75but lower than the superfocus micromixer of Hardt and Schonfeld.68
The post micromixers studied here will be well suited for
high-throughput applications. In contrast to the additional
secondary vortices used by most mixers, these mixers rely
on the primary vortices. Given the length scales of mixing
indicated in the previous graphs, a distributive feed over a
millimeter length scale (each substream size should be of the
order of the transverse post spacing or smaller) and an
increased mixer length, will substantially improve the performance of the post micromixers. This is indeed the case as
shown in Figure 12.
Conclusions
In this article, ow-driven instabilities at the microscale
were explored using 2-D CFD simulations. Flow past a single,
cylindrical post at the microscale was found to exhibit the same
nonlinear behavior as that at the macroscale and to be associated with symmetry breaking. Chemical reactions were found
to have an effect on oscillatory instabilities only when a change
in the uid density occurs. Such a density change may arise
when the number of moles is not conserved or under nonisothermal conditions. For isothermal ows the effect of chemistry on oscillations is weak. On the other hand, the observed
reactivity was found to be strongly affected by ow instabilities. As an example, increased conversions were found for
volumetric reactions as a result of enhanced mixing, and reduced conversions were observed for nonproduct-inhibited
positive order surface reactions arising from an effective reduction of the mass transfer coefcient, compared to a symmetric ow, caused by the fast removal of reactants from the
back of a post.
Aside from the ow chemistry interplay, ow-driven oscillations were found to enhance mixing at the microscale, above
a critical Reynolds number, with a relatively small pressure
drop. The performance of the micromixer was quantied in
terms of a mixing parameter at the exit. It appears that inter row
spacing can be tuned to enhance the mixing, and staggering
affects the variance in the exit more than the net mixing.
Potential improvements were suggested. However, further
studies are needed to fully optimize these mixers. Given that
posts are often fabricated for other reasons, such as increased
surface area, one could take advantage of them to enhance
mixing by working in the proper Re number regime. Advantages of this new method of mixing include the ease of fabrication and handling of particulate solutions compared to most
existing micromixer designs, the ease of scalability, and the
ability to handle relatively fast ows.
3202
December 2005
Acknowledgments
This work was supported by the Army Research Ofce under contract
DAAD19-01-1-0582. Any opinions, ndings, and conclusions or recommendations expressed are those of the authors and do not necessarily reect
the views of the Army Research Ofce.
Literature Cited
1. Jensen KF. Microreaction engineeringIs small better? Chem Eng
Sci. 2001;56:293-303.
2. Ehrfeld W. Design guidelines and manufacturing methods for microreaction devices. Chimia. 2002;56:598-604.
3. Schouten JC, Rebrov EV, de Croon MHJM. Miniaturization of heterogeneous catalytic reactors: Prospects for new developments in catalysis and process engineering. Chimia. 2002;56:627-635.
4. Schwalbe T, Autze V, Wille G. Chemical synthesis in microreactors.
Chimia. 2002;56:636-646.
5. Kursawe A, Dietzsch E, Kah S, Honicke D, Fichtner M, Schubert K,
Weimeier G. Selective reactions in microchannel reactors. Proceedings of the Third International Conference on Microreaction Technology. Berlin/New York: Springer-Verlag; 1999:213-223.
6. Kenis PJA, Ismagilov RF, Whitesides GM. Microfabrication inside
capillaries using multiphase laminar ow patterning. Science. 1999;
285:83-85.
7. Min J, Kim J-H, Kim S. Microuidic device for bioanalytical systems.
Biotechnol Bioprocess Eng. 2004;9:100-106.
8. Yun K-S, Yoon E. Microuidic components and bio-reactors for
miniaturized bio-chip applications. Biotechnol Bioprocess Eng. 2004;
9:86-92.
9. Makhijani VB, Raghavan J, Przekwas A, Przekwas AJ. Simulation of
biochemical reaction kinetics in microuidic systems. Proceedings of
the Third International Conference on Microreaction Technology.
Berlin/New York: Springer-Verlag; 1999:441-450.
10. Gunther A, Khan SA, Thalmann M, Trachsel F, Jensen KF. Transport
and reaction in microscale segmented gasliquid ow. Lab Chip.
2004;4:278-286.
11. Przekwas A, Wang D, Makhijani VB, Przekwas AJ. Microuidic
ltration chip for DNA extraction and concentration. Proceedings of
the Third International Conference on Microreaction Technology.
Berlin/New York: Springer-Verlag; 1999:488-497.
12. Fichtner M, Mayer J, Wolf D, Schubert K. Microstructured rhodium
catalysts for partial oxidation of methane to syngas. Ind Eng Chem
Res. 2001;40:3475-3483.
13. Trubac RE, Dautzenberg FM, Grifn TA, Paikert B, Schmidt VR,
Overbeek RA. Micro-engineered catalyst systems: ABBs advancement in structured catalytic packings. Catal Today. 2001;69:17-24.
14. Lyubovsky M, Karim H, Menacherry P, Boorse S, LaPierre R, Pfefferle WC, Roychoudhury S. Complete and partial catalytic oxidation
of methane over substrates with enhanced transport properties. Catal
Today. 2003;83:183-197.
15. Losey MW, Jackman RJ, Firebaugh SL, Schmidt MA, Jensen KF.
Design and fabrication of microuidic devices for multiphase mixing
and reaction. J Microelectromech Syst. 2002;11:709-717.
16. Ganley JC, Seebauer EG, Masel RI. Porous anodic alumina microreactors for production of hydrogen from ammonia. AIChE J. 2004;50:
829-834.
17. Norton DG, Wetzel ED, Vlachos DG. Fabrication of single-channel
AIChE Journal
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
43.
AIChE Journal
December 2005
44.
45.
46.
47.
48.
49.
50.
51.
52.
53.
54.
55.
56.
57.
58.
59.
60.
61.
62.
63.
64.
65.
66.
67.
68.
69.
70.
71.
3203
3204
December 2005
AIChE Journal
Imperfect initiator mixing greatly affects the stability and efciency of low density
polyethylene (LDPE) autoclave reactors. A combined simulation technique utilizing
compartment models and computational uid dynamics extends previous work in the
literature by providing a physically detailed picture of imperfect mixing. Analysis indicates that the effective volume for chain propagation in the autoclave reactor can expand
and contract in a continuous fashion as operating conditions change. As mixing becomes
poor, the effective reactive volume decreases, causing a reduction in initiator efciency,
but an expansion in the stable operation region. Examples demonstrate that accurate
prediction of the effective reaction volume is crucial for predicting LDPE autoclave
reactor behavior. A new mixing model that represents the feed plume by a series of
interconnected tanks with geometrically increasing volumes provides a favorable tradeoff
between accuracy and model complexity. 2005 American Institute of Chemical Engineers
AIChE J, 51: 32053218, 2005
Introduction
Low-density polyethylene (LDPE) accounts for about a
quarter of U.S. polyethylene production, and is an important
commodity polymer used in plastic bags, lms and coatings.1
The physics involved in the production of LDPE can be quite
complex, and the reaction kinetics, polymer properties, and
reactor stability in LDPE processes have, therefore, been studied for over 50 years.23
LDPE processes employ two types of reactors: (1) tubular
reactors, or (2) stirred autoclave reactors, and each type of
process is used extensively for LDPE production.4 Each of
these processes operates at high pressures (1,500 3,000 atmospheres), and with reactor exit temperatures ranging from about
200 to 300C.
In this work, we model a single compact type autoclave
reactor, such as the DuPont type (c.f. Figure 1). Typical processes have reactor volumes ranging between 250 and 2,000 L
and residence times on the order of 10 s to one min. The heat
Correspondence concerning this article should be addressed to W. H. Ray at
ray@engr.wisc.edu.
AIChE Journal
December 2005
generated by polymerization is quite high, and causes a temperature rise of about 12C13C for each percent of monomer
conversion.5 Due to relatively short residence times and thick
reactor walls, the capability of removing heat through wall
cooling is quite low.6 Thus, autoclave reactors can be considered adiabatic, and the exit temperature is attenuated by feeding the monomer in a cold condition.7 Due to safety limits on
the maximum temperature rise and the low residence times,
single pass conversion in autoclave reactors is usually below
20%. Despite this relatively low conversion to polymer, stability issues can be a major operational concern.
At temperatures above about 300C, a violently exothermic
reaction can decompose ethylene, leading to localized dark
spots in the polymer or global reactor runaway. Such ethylene
decomposition reactions are common in LDPE plant operation,
and occurrences of this reaction mechanism are known in the
industry as a decomp. Decomp episodes can be quite costly
due to environmental and safety concerns, and can ultimately
lead to signicant process downtimes. LDPE autoclaves can
also lose stability and migrate away from the desirable operating branch if internal temperatures drop too low and ignition
Vol. 51, No. 12
3205
December 2005
ment. The major parameter of the model was the recycle rate
from the large zone to the feed compartments. This recycle rate
parameter was chosen to match common reactor measurements. Using this three-compartment model, Marini and Georgakis1213 computed conversion and polymer properties. They
also computed the largest real eigenvalue for the reactor model,
and found that the range of operation over which this eigenvalue was negative was expanded with imperfect mixing.
Villa et al.14 carried out a full stability analysis of the LDPE
autoclave including the effects of the decomp reactions. They
showed that the perfectly mixed reactor has a tiny (or sometimes no) range of operating conditions for stable operation
where polymer is produced. However, by using simple threecompartment mixing models, they predicted a broad range of
productive, stable operation when there is imperfect mixing.
Villa et al.14 provide a clear physical explanation for the
stabilizing effect of imperfect mixing. Their study demonstrated that small regions in the feed plume of the reactor
signicantly reduce initiator and radical concentrations that
reach larger downstream zones. This causes a net reduction in
initiator efciency when compared to perfect mixing, and consequently reduces heat release and reactor sensitivity to temperature. Villa et al. showed that this reduction in temperature
sensitivity is analogous to a stabilizing controller that reduces
feed initiator concentration to the downstream zones when
operating temperature increases. Subsequent studies of imperfect mixing in the LDPE autoclave have also assessed aspects
of reactor operation and stability.8,15
Other research has examined more complex, and presumably
more realistic, mixing models in the context of process efciency and product properties.16 20 These studies have utilized
compartment mixing models of varying levels of complexity
and backow characteristics. Such models have been used to
predict industrial plant data in certain operating regions, but
little guidance exists concerning the appropriate level of detail
required to accurately capture the effects of mixing on LDPE
reactor operation.
Recently, computational uid dynamics (CFD) has been
used to provide an increased level of mixing detail in models of
autoclave and tubular LDPE reactors.2126 Such CFD studies
showed that mixing in autoclaves can play a very signicant
role in regions near initiator feeds, where large gradients in
temperature and composition exist.2124 These gradients can
subsequently affect process productivity and efciency. However, due to limitations in computing power, such detailed CFD
models have not been used to examine mixing effects on LDPE
autoclave operation over wide ranges in feed conditions or
design parameters. Furthermore, stability and polymer property
issues have not been studied thus far with the level of mixing
detail provided by CFD. It is envisioned that CFD will aid in
understanding the underlying physics that govern mixing effects on the operation of the LDPE autoclave.
This article applies a modeling approach that combines the
computational simplicity of compartment models with the mixing detail of CFD simulation. The approach models detailed
mixing effects on stability and productivity in the adiabatic
LDPE autoclave over wide operating ranges. General effects
governing spatial variations and stability are rst established in
the relatively simple three-compartment model used by Villa et
al.14 and then these simple models are compared to the more
detailed approach involving CFD. The analysis provides guideVol. 51, No. 12
AIChE Journal
Initiation:
feff kd
2R decomposition of initiator)
IO
fast
P1
R MO
Chain Propagation:
kp
Pn1
Pn MO
Termination by combination:
Quantity
Value
Inlet ow velocity
Residence time
Impeller rotation rate
Reactor volume
Mixture density
Heat capacity
Species diffusivity
Heat conductivity
Mixture viscosity
Initiator efciency
Reactor pressure
Heat of polymerization
Heat of decomposition
50 m/s
32.8 s
200 RPM
500 L
499 g/L
2.768 J/g-K
2.88 10-5 m2/s
0.1998 J/m-s-K
1.6 103 kg/m-s
1
2000 atm
21386 cal/mol
30200 cal/mol
*Note: Values for physical constants are taken from Read et al.22
ktc
Dnm
Pn Pm O
Ethylene decomposition:*
Cmon1.89kmd1 kpd1 0.0714kpd2
M O
decomp products
30200 cal/mol
*Note: Ethylene decomposition (decomp) is modeled as a heat release only;
decomp products are not explicitly tracked. This simplied form of the decomp
model proposed by Zhang et al.27 relates the rate of heat generation from
monomer decomp to lumped kinetic parameters (k md, k pd1, k pd2) from the
decomposition mechanism as follows28:
2
R heat,decomp HdecompCmon
1.89kmd kpd1 0.0714kpd2Cmon
Reaction Kinetics
Table 1 presents the reactions comprising the kinetic mechanism used in this work for modeling free-radical ethylene
homopolymerization in the high-pressure autoclave reactor. As
can be seen, the common free radical initiation, propagation,
and termination steps are modeled. In this work, a single
chemical initiator is used, and termination is presumed to occur
primarily by coupling in accordance with experimental evidence.29
Table 2 presents the rate constant parameters used in this
study to dene reactive source terms. The rate constants are
assumed to be of the Arrhenius form with both temperature and
pressure dependencies. A simplied form of the decomp model
Units
k0
E A (cal/mol)
V A (cal/atm-mol)
Source
Propagation (k p )
Initiator decomposition (k d )
Termination by coupling (k tc )
Initial monomer decomp. (k md )
Decomp. propagation 1 (k pd1 )
Decomp. propagation 2 (k pd2 )
L/(mol-s)
s1
L/(mol-s)
L/(mol-s)
L/(mol-s)
s1
1.14 107
1.06 1016
3.00 109
4.00 1019
1.59 1020
4.39 1020
7091
35560
2400
65000
65000
65000
0.477
0.0605
0.3147
0.1937
0.3218
0.1937
30
4
30
27
27
31
*Rate constants (k) are computed using the formula: k k 0 exp[(E A V A P)/R gasT], where P is the absolute pressure in atm, T is the absolute temperature
in K and R gas is the ideal gas constant.
AIChE Journal
December 2005
3207
feed fractions because the feed stream itself has enough energy
to support ignition of the entire reactor volume.
At the lower feed temperatures of 360 K and 300 K (Figure
3c d), imperfect mixing provides a desirable stable operating
branch, but no such stable conditions exist for the perfectlymixed model using the TBPOA initiator at the current residence time. For these lower feed temperatures, a signicant
temperature rise in the partially isolated feed zone must occur
in tandem with backmixing of hot material to decompose the
initiator and ignite the reactor. These conditions are provided
by the three-compartment imperfect mixing model, but not by
the perfectly-mixed model.
It should be noted that low feed temperatures in the range of
about 293363 K (20 90C) are common industrially,33 because they allow for greater adiabatic temperature rise and
higher monomer conversion before the dangerous decompositions reactions become important. Thus, the results of this section
emphasize the necessity of including mixing effects in models of
industrial LDPE autoclave reactors, since perfectly-mixed models
may not predict any desirable stable branch for industrially relevant values of feed temperature and residence time.
December 2005
AIChE Journal
Figure 4(a d). Effect of recycle ratio on spatial variations of temperature and composition.
An adiabatic, 3-compartment model is used with feed conditions of 300 K and 50 ppm of the TBPOA initiator.
December 2005
internal ows. Therefore, design parameters such as feed initiator fraction, feed temperature, and residence time are the
focus of highly-detailed mixing model studies in the subse-
Figure 5(a b). Effect of recycle ratio on fractional conversion and initiator consumption in each
compartment.
The feedstream is introduced at 300 K with 50 ppm of
the TBPOA initiator.
3209
Figure 6. Procedure for selecting and using a compartment model of the LDPE autoclave reactor.
December 2005
AIChE Journal
V HR
nzones
f rV
p,i
Rp,max
i1
(2)
n zones
i1
10 6 w I,f
x p MW I
(3)
December 2005
3211
reactor that intersects that feed and exit pipes, as well as the
impeller and stirring blades. From continuation and stability
analysis of the compartment model, the region of stable operation is also determined.
As in the studies of van der Molen et al.,9 10 the current
analysis shows that the initiator consumption has a minimum at
an intermediate internal operating temperature. This behavior
can be understood by examining the two radical contours. At
the higher amount of initiator in the feed (120 ppm), the total
polymer production is also higher, and the adiabatic reactor
exit temperature is about 535 K. At this temperature, the
initiator half-life is short compared to ow timescales, and the
initiator is poorly dispersed. This poor initiator dispersion can
be seen from the radical concentration prole shown for the
high initiator feed fraction. Due to the poor initiator dispersion,
the radicals are concentrated near the feed region, and the
bimolecular termination reaction destroys many radicals before
they can add signicant monomer to form polymer. Also, only
about 1/4 of the reaction volume is effectively used in the
polymerization reaction. This causes the initiator consumption
per kilogram polymer produced to be quite high, increasing
production costs.
Conversely, at a much lower initiator feed fraction of 30
ppm, the operating temperature drops considerably to about
500 K. At this temperature, the initiator half-life is similar to
ow time scales, and initiator is dispersed relatively evenly. As
a consequence, radicals are present in most of the reactor, the
bimolecular termination reaction is less dominant, and the
reactor volume is used effectively in polymerization. This
behavior is consistent with the better initiator usage, meaning
that less initiator is used per kilogram of polymer produced.
Unfortunately, as can be seen in Figure 7, the conditions for the
best radical dispersion and lowest specic initiator consumption are also near the reactor extinction limit. As a result,
reactor operation requires a tradeoff between efciency of
initiator usage and robust process stability.
December 2005
AIChE Journal
Figure 11. Effect of residence time on reactor exit temperature and stability range.
T f 360 K, w I,f 60 ppm TBPOA. The effective reactor
volume fraction is shown for 0.05. The base CFD case
for generating the 100-compartment model involves a feed
stream at 420 K with 120 ppm of TBPOA.
December 2005
3213
les, which show only a very small volume being used for
reaction at the highest residence time (328 s).
December 2005
to the three-compartment model that retain the volume variation of highly detailed models but do not require CFD simulation. A proposed modeling strategy to achieve this goal is
illustrated in Figure 13.
The simplied model presented in Figure 13 consists of a
sequence of perfectly-mixed reactors with recycle much like
the three-compartment model technique. However, many more
volumes are included to represent a plume that can shrink or
grow as the initiator half-life changes. Furthermore, the tanks
in the sequence increase geometrically in size as they move
away from the main feed, allowing for a wide range of effective reaction volumes. This modeling strategy is designed to
physically represent a feed plume that engulfs additional uid
as it progresses down the reactor.
The model illustrated in Figure 13 is similar in its physical
representation of the ow to the plume engulfment model
proposed by Zwietering.40 The plume engulfment model has
been applied to general reactive mixing problems, and considers incoming uid volumes to have varying ages. Fluid
volumes that have just entered the reactor are young, and
become closer to the bulk conditions as their age increases.
Unfortunately, the method of Zwietering requires the estimation of age time constants that may be difcult to obtain.
In contrast to the Zwietering mixing model,40 the plume
model presented in Figure 13 involves a more convenient
stirred-tank representation and is relatively simple to formulate. The plume model has only three parameters: the number
of compartments N compt. the recycle ratio R, and the factor
which describes the geometric rate of increase of compartment
Vol. 51, No. 12
AIChE Journal
volumes along the reactor train. The rate of side feed to the
reactors from the recycle stream is proportional to the volume
of the receiving compartment. This new simplied model could
be parameterized by knowledge of plant data, similar to the
three-compartment model approach. However, the plume
model also has the ability to expand and contract its physical
region of reaction nearly continuously and automatically as
conditions change, similar to the 100 compartment/CFD methodology.
The plume model presented here represents a straightforward approach that can be applied in practice, and it is expected to capture the stability region more completely than the
simpler three-compartment model. This hypothesis is conrmed in Figure 14, which compares the performance of the
plume model (for 1.2, R 15, and 20 zones) to that of the
three-compartment model, the 100 compartment model, and
CFD data. As can be seen from this gure, the plume model
captures a span in the stable region that is similar to the
100-compartment model. In general, the plume model performs
quite well given its simplicity, but it does not provide the same
level of spatial detail as the 100-compartment model. Other
advantages and disadvantages of the various modeling approaches employed here are summarized in Table 4, and should
be considered when modeling the operation of LDPE autoclave
reactors.
In general, because of its relative simplicity, good accuracy,
and ease of parameterization, the new plume model may represent a good compromise between effort and accuracy when
simulating the LDPE autoclave reactor. The plume model may
be especially attractive if a CFD model has not yet been
developed for the reactor, since model setup is often the most
time-consuming part of the overall CFD simulation process.41
December 2005
3215
Acknowledgments
The authors are indebted to the industrial sponsors of the University of
Wisconsin Polymer Reaction Engineering Laboratory, and to the U.S.
Department of Energy for nancial assistance. We also thank Iasson
Mustakis and Carlos Villa for their insight and support.
Notation
CFD
Cmon
Cp
Dn
EA
feff
fji
frV
fvol,i
H
k0
kd
kmd
kp
kpd1
kpd2
ktc
LDPE
M
MWI
MWmon
Ncompt
nzones
P
Pn
Qn
Rgas
R
R
3216
December 2005
RPM
Rheat,decomp
Rp
T
Tk
Tf
TBPOA
VA
Vtot
Vn
wI,f
wIk
wkp
xp
I
Hpoly
Hdecomp
0
k0
Literature Cited
1. McCoy M, Reisch MS, Tullo AH, Short PL, Thayer AM, Tremblay JF,
Storck WJ. Facts and gures for the chemical industry. Chem and Eng
News. 2003;81:25 66.
2. Roedel MJ. The molecular structure of polyethylene. I. Chain branching in polyethylene during polymerization. J A Chem Soc. 1953;75:
6110 6112.
3. Kiparissides C, Verros CG, MacGregor JF. Mathematical modeling,
optimization, and quality control of high-pressure ethylene polymerization reactors. J Macromolecular Sci. Reviews in Macromolecular
Chemistry and Physics. 1993;C33:437527.
4. Doak KW. Low density polyethylene. In: Mark HF. Encyclopedia of
Polymer Science and Technology. New York: Wiley Interscience
Publishing, 1985:383 429.
5. Beasley JK. Polymerization at High Pressure. In: Allen SG, Bevington
JC. Comprehensive Polymer Science. New York: Pergammon Press,
1989:273282.
6. Gemassmer AM. Autoclave Process for the High Pressure Polymerization of Ethylene. Erdol und Kohle-ErdgasPetrochemie. 1978;31:
221228.
7. Christl RJ, Roedel MJ. USA Patent No. 2897183, 1959.
8. Pladis P, Kiparissides C. Steady-State Multiplicity in High Pressure
Ethylene Polymerization Reactors. 3rd Annual AIChE Polymer Producers Conference, Houston, TX: 1999.
9. van der Molen TJ, van Heerden C. The Effect of Imperfect Mixing on
the Initiator Productivity in the High Pressure Radical Polymerization
of Ethylene. 1st International Symposium on Chemical Reaction Engineering, Washington, D.C. American Chemical Society; 1970.
10. van der Molen TJ, Koenen A, Oosterwijk H, van der Bend H. Lightoff temperature and consumption of 16 initiators in LDPE production.
Quaderni dellIngegnere Chimico Italiano. 1982;18:715.
11. Georgakis C, Marini L. The Effect of Mixing on Steady-State and
Stability Characteristics of Low Density Polyethylene Vessel Reactors. Seventh International Symposium on Chemical Engineering,
Boston, Massachusetts, American Chemical Society; 1982.
12. Marini L, Georgakis C. Low-Density Polyethylene Vessel Reactors.
Part I: Steady State and Dynamic Modeling. AIChE J. 1984;30:401
408.
13. Marini L, Georgakis C. Low-Density Polyethylene Vessel Reactors.
Part II: A Novel Controller. AIChE J. 1984;30:409 414.
14. Villa CM, Dihora JO, Ray WH. Effects of Imperfect Mixing on
Low-Density Polyethylene Reactor Dynamics. AIChE J. 1998;44:
1646 1656.
15. Lee HJ, Yeo YK. Modeling and Simulation of High Pressure Autoclave Polyethylene Reactor Including Decomposition Phenomena. J
Chem Eng of Japan. 2000;33:323329.
AIChE Journal
16. Donati G, Gramondo M, Langianni E, Marini L. Low density polyethylene in vessel reactors. Quaderni dellIngegnere Chimico Italiano.
1981;17:88 96.
17. Singstad P. Modelling and Multivariable Control of High Pressure
Autoclave Reactors for Polymerization of Ethene, The Norwegian
Institute of Technology; 1992. PhD Diss.
18. Chan WM, Gloor PE, Hamielec AE. A Kinetic Model for Olen
Polymerization in High-Pressure Autoclave Reactors. AIChE J. 1993;
39:111126.
19. Nordhus H, Moen , Singstad P. Prediction of Molecular Weight
Distribution and Long-Chain Branching Distribution of Low-Density
Polyethylene from a Kinetic Model. J Macromol Sci Pure Appl Chem.
1997;A34:10171043.
20. Ghiass M, Hutchinson RA. Simulation of Free Radical High-Pressure
Copolymerization in a Multizone Autoclave Reactor: Compartment
Model Investigation. Macromol Symposia. 2004;206:443 456.
21. Torvik R, Gravdahl AR, Fredriksen GR, Moen , Laurell J. Design of
HPPE stirred autoclaves using 3D computational uid dynamics. 5th
International Workshop on Polymer Reaction Engineering: Berlin;
1995.
22. Read NK, Zhang SX, Ray WH. Simulations of a LDPE Reactor Using
Computational Fluid Dynamics. AIChE J. 1997;43:104 117.
23. Tosun G, Bakker A. A study of macrosegregation in low-density
polyethylene autoclave reactors by computational uid dynamic modeling. Ind Eng Chem Res. 1997;36:296 305.
24. Zhou W, Marshall EM, Oshinowo L. Modeling LDPE tubular and
autoclave reactors. Ind Eng Chem Res. 2001;40:55335542.
25. Fox RO. Computational methods for turbulent reacting ows in the
chemical process industry. Revue de LInstitut Francais du Petrole.
1996;50(2):215243.
26. Kolhapure NH, Fox RO. CFD analysis of micromixing effects on
polymerization in tubular low-density polyethylene reactors. Chem
Eng Sci. 1999;54:32333242.
27. Zhang SX, Read NK, Ray WH. Runaway phenomena in low-density
polyethylene autoclave reactors. AIChE J. 1996;42:29112925.
28. Wells GJ. Modeling methods for predicting polymer process dynamics
and product properties: the Effects of Mixing and Catalyst Characteristics, University of Wisconsin-Madison; 2003. PhD Diss.
29. Odian G. Principles of Polymerization. New York: John Wiley and
Sons; 1991.
30. Chen CH, Vermeychuk JG, Howell JA, Ehrlich P. Computer Model
for Tubular High-Pressure Polyethylene Reactors. AIChE J. 1976;22:
463 471.
31. Zhang SX. Modelling and Experimental Studies of Free Radical
Polymerization Reactors. University of Wisconsin-Madison; 1996.
PhD Diss.
32. Zhang SX, Ray WH. Modeling of Imperfect Mixing and Its Effects on
Polymer Properties. AIChE J. 1997;43:12651277.
33. Sittig M. Low Density Polyethylene Processes. Polyolen Production
Processes: Latest Developments. Noyes Data Corporation; 1976.
34. Kwag BG, Choi KY. Effect of Initiator Characteristics on HighPressure Ethylene Polymerization in Autoclave Reactors. Ind Eng
Chem Res. 1994;33:211217.
35. Goto S, Yamamoto K, Furui S, Sugimoto M. Computer Model for
Commercial High-Pressure Polyethylene Reactor Based on Elementary Reaction Rates Obtained Experimentally. J Appl Poly Sci: Applied
Polymer Symposium. 1981;36:21 40.
36. Smit L. The Use of Micromixing Calculations in LDPE-Reactor Modelling. 4th International Workshop on Polymer Reaction Engineering;
1992.
37. Wells GJ, Ray WH. Methodology for modeling detailed imperfect
mixing effects in complex reactors. AIChE J. 2005;51:1508.
38. Fluent Incorporated. Fluent Users Guide: Version 6.0; 2001.
39. Britton LG, Taylor DA, Wobser DC. Thermal Stability of Ethylene at
Elevated Pressures. Plant/Operations Progress. 1986;5:238 251.
40. Zwietering TN. A backmixing model describing micromixing in single-phase continuous-ow systems. Chem Eng Sci. 1984;39:1765
1778.
41. Versteeg HK, Malalasekera W. An Introduction to Computational
Fluid Dynamics: The Finite Volume Approach. London: Longman
Group; 1995.
AIChE Journal
December 2005
Appendix
State equations
The state equations for zone k of a compartment model with
any number of compartments are given in the equations below.
If not specied, all rate constants, concentrations, mass fractions, and temperatures are at the conditions of the compartment k. Compartment volumes are assumed to be time invariant, and constant physical properties are also assumed.
dw Ik
1
dt
Vk
d 0k
1
dt
Vk
k
n
n
I
(A1)
n0
f Q 2k f C k
n
k
n
n
0
d eff
tc
(A2)
n0
f Q w k C
Ncompt
k
n
n
p
0 MWmon
mon
n0
f Q T H
Ncompt
k
n
Ncompt
dw pk
1
dt
Vk
1
dT k
dt
Vk
f Q w k C MW
Ncompt
k Cmon0
Cp
poly p
n0
(A3)
2
Hdecomp[Cmon
(1.89kmd kpd1 ) 0.0714kpd2 Cmon]
Cp
(A4)
100-Compartment models
The 100-compartment model used in this work is selected
from a CFD base case using criteria that minimize variations in
the temperature, propagation rate, and termination rate in each
selected zone. Details of this selection process are given in
Wells and Ray.37
3-Compartment model
For the 3-compartment model (see Figure 2), the volumes
are taken to be the values used by Zhang and Ray,32 namely
V 1 V tot/24
V 2 2V tot/24
V 3 21V tot/24
(A5)
3217
f 12 1
V1
f 23 1
V tot
N compt1
i0
R
f
31 R
1
3
f 32
V i V i1; 1 i N compt
2R
31 R
f ii 1, i 1 . . . 3
f ij 0, all other cases
(A6)
f 01 1
f ii1 1; 1 i N compt
Plume model
The plume model (see Figure 13) denes the ow fractions
and volume fractions based on the number of compartments in
the model (N compt), the recycle ratio (R), and the volume
growth factor (). Based on the denitions given in Figure 13,
the compartment volumes can be related to the total reactor
volume and growth factor using the following recursive relationship
3218
(A7)
December 2005
f ii 1, i 1 . . . Ncompt
i
f Ncompt
RVi
; 1 i Ncompt
R 1Vtot VNcompt
(A8)
Manuscript received Nov. 24, 2004, and revision received Mar. 10, 2005.
AIChE Journal
Introduction
Dynamic models for heterogeneous reactors are composed
of distributed parameter, ordinary, and partial differential equa-
AIChE Journal
December 2005
3219
December 2005
linear parabolic reaction diffusion equations, which is the spatial integral averaging of the dynamic reaction rate. In this
form, it is shown that the DEF becomes the transfer function of
a model governing the dynamics of the catalyst particle reaction rate. However, the resulting transfer function is innitedimensional. This implies that, contrary to the steady-state
case, an exact simple (such as nite-dimensional) modeling of
the reaction rate dynamics is not possible. To overcome this
problem, an approach to obtain low-order models based on an
approximate model matching is proposed. For a rst-order
model, the time constant is computed showing its dependency
with the Thiele modulus. In fact, the results show that the
(dominant) time constant is a decreasing function of the Thiele
modulus. In this form, the faster the chemical reaction dynamics, the faster the reaction diffusion catalyst particle dynamics.
Numerical simulations are used to illustrate our ndings.
Classical EF concept
As mentioned earlier, the classical EF concept pertains to
nondynamic operating conditions. To recall the classical EF
concept, consider an isothermic catalyst particle where a single
rst-order chemical reaction is taking place. In a general geometry, by assuming constant diffusivity D and a Fickian
diffusion mechanism J Dc, the nondynamic process can
be modeled as follows (throughout, overbars denote nondynamic conditions):
Dc kc 0
(1)
2
2
2
x 12 x 22 x 32
for all x D
for all x D
(2)
AIChE Journal
is, c (x) c b for all x D]. The EF expresses the reaction rate
in a catalyst particle by the rate under surface/bulk conditions.1
That is, the classical nondynamic EF (denoted by ) is dened
as follows:
def
D kc dV
D kc bdV
(3)
kL
D
def
(4)
tanh
2I1
I0
1
1
2
3
tanh
def
def
R b
kc bdV
: R b 3 R
Rs
kC x, sdV
(7)
kc xdV
(6)
(8)
C x, s C bs
for all x D
Cx, s is finite
for all x D
and
AIChE Journal
December 2005
(9)
3221
d D/L 2
be the diffusion characteristic frequency. Introduce the normalized Laplace variable as (recall that s has frequency units)
s : R bs 3 Rs
(10)
sn
(14)
(15)
s n,
tanhsn
sn
2I1 sn
sn I0 sn
1
1
3
sn tanhsn sn 2
(11)
s L 2
D
(12)
s s
(13)
s
d
December 2005
AIChE Journal
December 2005
3223
reaction diffusion process is not seriously deteriorated. Disturbances with frequency components in the range d do
not have a signicant effect on the supply of molecules to the
reaction mechanism.
(2) A high-frequency regime where the performance is seriously depleted. Because of high-frequency perturbations, the
molecules are unable to penetrate the particle, such that the
reaction rate dynamic R(sn) is seriously affected.
It is noticed that for the low-frequency regime (n, )
(). That is, the DEF is almost equal to the nondynamic EF.
This shows that for long times (corresponding to n 3 0) or for
low-frequency perturbations one can take the quasi-steadystate model R(t) ()Vkcb(t) to compute the reaction rate
behavior R(t). Indeed, this is an assumption commonly made in
practice.
(17)
where
def
gs n, kV s n,
Equation 17 describes the dynamics R(sn) affected by the
boundary concentration dynamics Cb(sn). Because the operator
g(sn, ) is causal, there exist matrices (A, B, C) such that the
dynamics of Eq. 17 can be described in state-space form. That
is, if z is a state vector, it is possible to write10
dzt n
Azt n Bc bt n
dt n
Rt n Czt n
(18)
s n,
(19)
hs n,
3224
December 2005
ns n 1
AIChE Journal
(20)
Imn ,
kV
ns n 1
(21)
n,
ns n 1
(22)
n, Ren , i Imn ,
(24)
f n,
ni n 1
This function can be written in terms of its real and imaginary parts as follows:
f n,
n n
i
n 2 n2 1
n 2 n2 1
(27)
Im0, n n n
where Im[(0, )] is the derivative of Im[(0, )] with
respect to the frequency n. By assuming the equality Im[(0,
)]n ()n()n, an estimate of the time constant is
given by
Im0,
(28)
(25)
(3) Pose an approximate model-matching problem to compute the time constant n(). By matching the real and imaginary parts of f (n, ) and (n, ) (see Eqs. 24 and 25), the
following two approximation problems should be solved:
AIChE Journal
n n
n 2 n2 1
(23)
(26)
and
n 2 n2 1
Ren ,
December 2005
Im0,
Im, Im0,
3225
n 2
dRt n
d 2Rt n
2 n
Rt n Vk
dt n
dt n2
(30)
1 n 2 n2 i2 n n
(31)
1 n 2 n2
2n n
i
n,
n ,
(32)
1 n 2 n2
n ,
(33)
and
Vk
2 2
n s n 2 n s n 1
(29)
December 2005
AIChE Journal
Imn ,
2n n
n ,
(34)
1
Re0, n2 2n 2 n2
2
(35)
1
Re0,
4
(36)
In this work, we have considered reaction diffusion processes with linear reaction rates. This allowed a model-reduction methodology based on a transfer function; that is, the
effectiveness factor in dynamic conditions. Although by itself
the DEF concept introduced in the preceding sections can be of
interest to researchers working in reaction engineering, the
model-reduction problem for linear reaction rates is rarely of
practical importance for such applications as process control.
In fact, it is very likely that the reaction rates are nonlinear in
real processes. As described earlier, the model-reduction technique based on frequency-response tools cannot be used for
reaction diffusion processes with nonlinear reaction rates because a transfer function cannot be computed (only local transfer functions can be computed). To overcome this problem,
Szukiewicz11 proposed to linearly approximate the reaction
rate and compute a rst-order model reduction based on such
approximation. By using heuristic arguments, the nonlinear
nature of the reaction diffusion processes is subsequently recovered. A drawback of Szukiewiczs proposal is its lack of a
systematic procedure to evaluate the model-reduction error.
Moreover, high-order extensions to improve the approximation
are not reported.
Well-established model-reduction methods (for example,
Galerkin) for reaction diffusion equations are based on the
following observation. The concentration dynamics is driven
by diffusion mechanisms modeled as Dc, and the reaction
activity is described, in general, by a nonlinear function f (c).
The latter is a pointwise function, and so it does not introduce
spatial memory. The former, given its distributed nature,
involves an innite-dimensional operator, which introduces an
innite number of dynamic modes. Thus, a procedure aimed at
obtaining a nite-dimensional model should focus on the approximation of the diffusion operator. This idea has found
sound applications in the computation of nite-dimensional
controllers for the reaction diffusion processes by Galerkin
methods.6,8,9 Given the linear nature of the diffusion operator,
the same idea can be exploited to obtain reduced-order models
by classical response-frequency tools. The rationale behind this
proposal is that, as in Galerkin methods, a reduced-order model
retains the dominant modes of the diffusion operator.
In this way, a rst-order model contains an estimate of the
dominant time constant of the process dynamics that, for many
process applications, sufces to compute an industrial [such as
proportional integral (PI)] feedback controller.12
To illustrate these ideas in a preliminary way, consider a
reaction diffusion process with a nonlinear reaction rate
December 2005
c
Dc fc
t
with Dirichlet-type boundary conditions (see Eq. 9). In the
spirit of the Galerkin method, let us compute an approximation
to the distributed part of the above equation. To this end, given
the pointwise nature of f (c), consider the reactionless equation
c/t Dc and compute the corresponding transfer function
Vol. 51, No. 12
3227
as done above. This is equivalent to taking (sn, 0). Subsequently, compute a rst-order approximation to give (0) 1
and n(0) 1/3 (see Figure 4). This implies that the dominant
time constant for the diffusion operator (in rectangular coordinates) is equal to 1/3. By reincorporating the nonlinear reaction
rate evaluated at average concentration conditions, the approximate rst-order model for the average concentration dynamics
c can be written as
dc 3D
c t c fc
dt
Rp b
where cb(t) is the boundary concentration and Rp is the particle
radius. It is interesting to remark that for spherical coordinates
the dominant time constant is n(0) 1/15, which recovers the
result reported by Szukiewicz.11 One advantage of the approach described above is that high-order approximations can
be obtained by increasing the approximation order of the transfer function (sn,0). In such a case, the models can be expressed as in the following:
dc
t fc
dt
where (t) is the output of a linear system and (s) h(s)cb(s),
and h(s) is a rational approximation to the transfer function
h(s) : cb(s) 3 c (s) corresponding to the reactionless diffusion
system.
It is interesting to note that when the method described
above is used in the linear case, the approximate dynamics are
given by
to obtain reduced-order models, the issue of computing reduced-order models for nonlinear kinetics should be explored
in a subsequent work.
Simulations
This section is used to illustrate the performance of the
model Eq. 23. For rectangular geometry, the dynamic reaction
diffusion problem can be written as
dc
3c bt c 2c
dt
c 2c
2c
t n x 2
December 2005
(37)
for x 1
(38)
for x 0
Rt n Q 2
c x, t ndx
(39)
(40)
AIChE Journal
December 2005
Figure 8. Frequency response of the exact and the approximate solution for 1.0 and three different values of the oscillation frequency.
Vol. 51, No. 12
3229
ences between the numerical results discussed above and numerical results from Szukiewiczs estimation were obtained.
Such differences are found in the high-frequency (such as
initial-conditions effect) regime.
Conclusions
In this paper, the problem of representing the distributed
parameter dynamics of a catalyst particle by means of loworder (lumped parameter) models was addressed. As a preliminary step, an extension of the classical effectiveness factor
(EF) concept to dynamic conditions was introduced. To do this,
the EF was interpreted as an operator that transforms total
reaction rates from surface to particle conditions. Based on this
interpretation, a dynamic EF (DEF) was dened as a linear
operator that transforms total reaction signals from surface to
particle conditions. In this way, the DEF retains some information on the performance of a particle under dynamic perturbations. For common geometries, the computation of the DEF
can be easily obtained from the nondynamic expression by a
direct substitution of the classical Thiele modulus by a sort of
complex Thiele modulus. The DEF results showed that the
better catalyst performance is found at steady-state conditions;
that is, dynamic perturbations deteriorate the catalyst particle
effectiveness.
By departing from the DEF concept, a systematic method to
obtain low-order models for the reaction dynamics was proposed. An algorithm to estimate the time constant for rstorder models was proposed by using model-matching methods.
The results showed that the larger the Thiele modulus, the
smaller the time constant. Although our analysis showed that a
stable second-order model cannot be computed from matching
3230
December 2005
Literature Cited
1. Aris R. The Mathematical Theory of Diffusion and Reaction in Permeable Catalysts. Vol. 1. Oxford, UK: Clarendon Press; 1975.
2. Kim DH. Linear driving force formulas for diffusion and reaction in
porous catalyst. AIChE J. 1989;35:343-346.
3. Goto M, Hirose T. Approximate rate equation for intraparticle diffusion with and without reaction. Chem Eng Sci. 1993;48:1912-1915.
4. Szukiewicz MK. New approximate model for diffusion and reaction in
a porous catalyst. AIChE J. 2000;46:661-665.
5. Christodes PD, Daoutidis P. Feedback control of hyperbolic PDE
systems. AIChE J. 1996;42:3063-3086.
6. Christodes PD, Daoutidis P. Finite-dimensional control of parabolic
PDE systems using approximate inertial manifolds. J Math Anal Appl.
1997;216:398-420.
7. Sheng P, Zhou MY. Dynamic permeability in porous media. Phys Rev
Lett. 1988;61:1591-1594.
8. Baker J, Christodes PD. Finite-dimensional approximation and control of nonlinear parabolic PDE systems. Int J Contr. 2000;73:439456.
9. Christodes PD. Nonlinear and Robust Control of Partial Differential
Equation Systems: Methods and Applications to Transport-Reaction
Processes. Boston, MA: Birkhauser; 2001.
10. Chen CT. Introduction to Linear System Theory. New York, NY: Holt,
Rinehart & Winston; 1970.
11. Szukiewicz MK. An approximate model for diffusion and reaction in
a porous pellet. Chem Eng Sci. 2002;57:1451-1457.
12. Morari M, Zariou E. Robust Process Control. New York, NY:
Prentice-Hall; 1989.
Manuscript received Jan. 11, 2005, and revision received Apr. 4, 2005.
AIChE Journal
Vemuri Balakotaiah
Dept. of Chemical Engineering, University of Houston, Houston, TX 77204
DOI 10.1002/aic.10574
Published online September 6, 2005 in Wiley InterScience (www.interscience.wiley.com).
A two-scale continuum model is developed to describe transport and reaction mechanisms in reactive dissolution of a porous medium, and used to study wormhole formation
during acid stimulation of carbonate cores. The model accounts for pore level physics by
coupling local pore-scale phenomena to macroscopic variables (Darcy velocity, pressure
and reactant cup-mixing concentration) through structure-property relationships (permeability-porosity, average pore size-porosity, and so on), and the dependence of mass
transfer and dispersion coefcients on evolving pore scale variables (average pore size
and local Reynolds and Schmidt numbers). The gradients in concentration at the pore
level caused by ow, species diffusion and chemical reaction are described using two
concentration variables and a local mass-transfer coefcient. Numerical simulations of
the model on a two-dimensional (2-D) domain show that the model captures the different
types of dissolution patterns observed in the experiments. A qualitative criterion for
wormhole formation is developed and it is given by O(1), where keffDeT/uo.
Here, keff is the effective volumetric dissolution rate constant, DeT is the transverse
dispersion coefcient, and uo is the injection velocity. The model is used to examine the
inuence of the level of dispersion, the heterogeneities present in the core, reaction
kinetics and mass transfer on wormhole formation. The model predictions are favorably
compared to laboratory data. 2005 American Institute of Chemical Engineers AIChE J, 51:
32313248, 2005
Introduction
Acid treatment of carbonate reservoirs is a widely practiced
oil and gas well stimulation technique. The primary objective
of this process is to increase the production rate by increasing
permeability of the damaged zone near the wellbore region.
The injected acid dissolves the material near the wellbore and
creates ow channels that establish a good connectivity beCorrespondence concerning this article should be addressed to V. Balakotaiah at
bala@uh.edu.
AIChE Journal
December 2005
3231
Literature Review
Figure 1. Typical dissolution patterns observed in carbonate acidizing: (a) face dissolution, Q 0.04
cc/min, PVinj 43.1, (b) conical Q 0.11 cc/
min, PVinj 10, (c) wormhole Q 1.05 cc/min,
PVinj 0.8, (d) ramied Q 10 cc/min, PVinj
2.1, and (e) uniform Q 60 cc/min, PVinj 6.7.
Hydrochloric acid is used in these experiments and the acid
injection rate is increased from (a) to (e) (Fredd and Fogler15).
The cores are approximately 3.8 cm in dia. and 10.2 cm in
length.
December 2005
Wormhole formation during reactive dissolution of carbonates is a subject that has been actively studied in the last thirty
years. To explain wormhole formation, numerous models,
ranging from detailed pore-scale models (for example, network
models) that account for reaction, transport and dissolution at
the pore scale to single wormhole (tube) models that consider
only the mechanisms occurring inside the wormholes, have
been developed in the literature. In this section, a brief review
of different models developed to study wormhole formation in
carbonates is presented.
Relating the important dimensionless groups of the system to
experimental observations is one of the approaches followed to
model acidization process. For example, Fredd and Fogler12,15
and Fredd16 have reported the dependence of wormhole formation on the Damkohler number and predicted an optimum
Damkohler number of 0.29 for different uid/mineral systems.
Daccord et al.7,8,17 also used a dimensionless parameter based
approach and coupled it with the concept of fractals to obtain
the propagation rate of wormholes. The use of a few dimensionless groups to explain experimental observations is a difcult exercise because the actual number of parameters in the
system is large and it is difcult to study wormholing phenomena systematically using this approach.
observed
in
The pore volume of acid (HCl, 22C) required for breakthrough is high at very low and very high injection rates, and
is minimum at optimum injection rate, Q opt. The length and
diameter of the cores are 10.2 cm and 3.8 cm, respectively.
The initial porosities and permeabilities are in the range of
0.15 0.2 and 0.8 2 md, respectively.
AIChE Journal
December 2005
3233
(1)
U0
t
(2)
C f
UCf De Cf kc a Cf Cs (3)
t
k cC f C s RC s
RC sa
t
s
(4)
Cs
(5)
December 2005
Cf
1
ks
kc
(6)
In the kinetically controlled regime (k s k c ), the concentration at the uid-solid interface is approximately equal to the
concentration of the uid phase (C s C f ). In the mass transfer
controlled regime (k s k c ), the value of concentration at the
uid-solid interface is very small (C s 0). Since the rate
constant is xed for a given acid, the magnitude of the ratio
k s /k c is determined by the local mass-transfer coefcient k c ,
which is a function of the pore geometry, the reaction rate and
the local hydrodynamics. Due to dissolution and heterogeneity
in the medium, the ratio k s /k c is not a constant in the medium
but varies with space and time which can lead to a situation
where different locations in the medium experience different
regimes of reaction. To describe such a situation it is essential
to account for both kinetic and mass transfer controlled regimes
in the model. Equation 5 describes the evolution of porosity in
the domain due to reaction.
To complete the model (Eqs. 15), information on the permeability tensor K, dispersion tensor De , mass-transfer coefcient k c , and interfacial area a is required. These quantities
depend on the pore structure and are inputs to the Darcy scale
model from the pore scale model. Instead of calculating these
quantities from a detailed pore scale model taking into consideration the actual pore structure, we use structure-property
relations that relate permeability, interfacial area and average
pore radius of the pore scale model to its porosity. However, if
a detailed calculation including the pore structure can be made,
then the quantities K, De , k c , and a obtained from such a
calculation can be used as inputs from the pore scale model to
the Darcy scale model.
AIChE Journal
semiempirical relations that relate the properties to local porosity. The relative increase in permeability, pore radius and
interfacial area with respect to their initial values are related to
porosity in the following manner
1 o
K
K o o o1
rp
ro
K o
K o
(7)
(8)
and
a r o
a o or p
(9)
December 2005
2k cr p
dh
Sh 0.35
Dm
x
0.5
Rep1/ 2 Sc1/3
(10)
(11)
3235
Udh
Dm
(12)
D eX
os XPe p
Dm
(13)
DT
D eT
os TPe p
Dm
(14)
and
1
D eX
3Pep
os Pe pln
Dm
6
2
Regime
December 2005
7 104
7 103
0.7
1.4 10
1.4 103
0.14
x
,
L
rp
,
ro
a
,
ao
y
,
L
z
,
L
cf
2k sr o
,
Dm
Da
U
,
uo
t
,
L/uo
K
,
Ko
(16)
Pe p
r
D eX
os Pe p2.
Dm
Face
Wormhole
Uniform
(15)
and
3236
k sa oL
,
uo
Cf
,
Co
N ac
Pe L
cs
Cs
,
Co
P Pe
u oL
Ko
Co
,
s
u oL
,
Dm
2r o
,
L
H
L
AIChE Journal
resented by K o , a o and r o , respectively. The parameters obtained after making the equations dimensionless are the Thiele
modulus 2, the Damkohler number Da, the acid capacity
number N ac , the axial Peclet number Pe L , aspect ratio o , and
. The Thiele modulus (2) is dened as the ratio of diffusion
time to reaction time based on the initial pore size and the
Damkohler number (Da) is dened as the ratio of convective
time to reaction time based on the length scale of the core. The
acid capacity number (N ac ) is dened as the volume of solid
dissolved per unit volume of the acid, and the axial Peclet
number Pe L is the ratio of axial diffusion time to convection
time. Notice that in the above parameters, inlet velocity u o
appears in two parameters Da and Pe L . To eliminate inlet
velocity from one of the parameters, so that the variable of
interest, that is, injection velocity, appears in only one dimensionless parameter (Da) we introduce a macroscopic Thiele
modulus 2 which is dened as 2 k s a o L 2 /D m DaPe L .
The macroscopic Thiele modulus is a core scale equivalent of
the pore scale Thiele modulus (2), and is independent of
injection velocity. The dimensionless equations in the twodimensional (2-D) case are given by
p
p
,
x
y
u,
(17)
0
t x y
DaA cf
c f uc f c f
t
x
y
x
2 r
1
Sh
X ur
cf
x
y
os Da
2
os Da
cf
T ur
2
y
DaN acA c f
t
2r
1
Sh
(18)
(19)
p
1@x0
x
p0@x1
(22)
p
0 @ y 0, 0
y
(23)
cf 1 @ x 0
(24)
c f
0@x1
x
AIChE Journal
(21)
(25)
December 2005
and y o
(26)
cf 0 @ t 0
(27)
x, y, a, l 0 f a,l @ t 0
(28)
(20)
c f
0@y0
y
PV FaceD
s1 o 1 o
C o o
N ac o
(29)
k sC oa
2r
s 1
Sh
(30)
3237
t bth
s
k sC o
2r
Sh
(31)
PV UniformD
t bthu o
su o
oL
k sC o a o oL
2r
Sh
1
DaN ac o
2r
Sh
December 2005
erogeneity is an important factor that promotes pattern formation during reactive dissolution. Without heterogeneity, the
reaction/dissolution fronts would be uniform despite an adverse
mobility ratio between the dissolved and undissolved media. In
a real porous medium, the presence of natural heterogeneities
triggers instability leading to different dissolution patterns. To
simulate these patterns numerically, it is necessary to introduce
heterogeneity into the model. Heterogeneity could be introduced in the model as a perturbation in concentration at the
inlet boundary of the domain or as a perturbation in the initial
porosity or permeability eld in the domain. In the present
model, heterogeneity is introduced into the domain as a random
uctuation of initial porosity values about the mean value of
porosity as given by Eq. 28. The two important parameters
dening heterogeneity are the magnitude of heterogeneity a,
and the dimensionless length scale l. The effect of these
parameters on wormhole formation is investigated in this, and
the following subsections.
The inuence of the magnitude of heterogeneity (a) is
studied by maintaining the length scale of heterogeneity constant (which is the grid size) and varying the magnitude from
a small to a large value. Figures 6a 6e show the porosity
proles of numerically simulated dissolution patterns for different Damkohler numbers on a domain with a large magnitude
of heterogeneity in initial porosity distribution. The uctuations (f) in porosity ( 0.2 f ) are uniformly distributed in
the interval [0.15, 0.15] (a 0.75). Figures 6f 6j show
the porosity proles for the same Damkohler numbers used in
Figures 6a 6e, but with a small magnitude of heterogeneity in
the initial porosity distribution. The uctuations (f) in porosity
( 0.2 f ) for this case are distributed in the interval
[0.05, 0.05] (a 0.25). It could be observed from the
gures that wormholes do not exhibit branching when the
magnitude of heterogeneity is decreased. This observation suggests that branching of wormholes observed in carbonate cores
Vol. 51, No. 12
AIChE Journal
Figure 6. Porosity proles at different Damkohler numbers with uctuations in initial porosity distribution in the interval [0.15, 0.15] are shown in
Figures a e.
Figures fj show porosity proles for the same Damkohler
numbers as used in Figures a e, but for uctuations in the
interval [0.05, 0.05]. The values of Damkohler numbers for
different patterns are: (a) Da 3 10 4 ( o 30), (b)
Da 10 4 ( o 10), (c) Da 500 ( o 0.5), (d) Da
40 ( o 0.04), and (e) Da 1 ( o 0.001). The values
of parameter o are reported for later use. The values of other
parameters xed in the model are 2 10 6 , 2 0.07,
N ac 0.1, o 0.4.
could be a result of a wide variation in magnitude of heterogeneities present in the core. Figures 6a 6j show that at very
large Damkohler numbers (low injection rates), the acid reacts
soon after it contacts the medium resulting in face dissolution,
and at low values of Damkohler number (high injection rates),
acid produces a uniform dissolution pattern. Wormholing patterns are created near intermediate/optimum values of the
Damkohler number. While changing the magnitude of heterogeneity changes the structure of the wormholes, an important
observation to be made here is that the type of dissolution
pattern (wormhole, conical, and so on) remains the same at a
given Damkohler number for different magnitudes of heterogeneity. Thus, heterogeneity is required to trigger the instability and its magnitude determines wormhole structure, but the
type of dissolution pattern formed is governed by the transport
and reaction mechanisms. Figure 7 shows the pore volume of
acid required to breakthrough the core at different injection
rates with different levels of heterogeneity for the porosity
proles shown in Figure 6. The curves show a minimum at
intermediate injection rates because of wormhole formation. It
could be observed from the breakthrough curves that the minimum pore volume/breakthrough time and optimum injection
rate (Damkohler number) are approximately the same for both
levels of heterogeneity.
The second parameter related to heterogeneity that is introduced in the model is the length scale of heterogeneity l. The
effect of this parameter on wormhole structure is dependent on
the relative magnitudes of convection, reaction and dispersion
levels in the system. The role of this parameter on wormhole
formation is thus discussed after investigating the effects of
convection, reaction and transverse dispersion in the system.
AIChE Journal
December 2005
(b) Convection and Transverse Dispersion. In the previous subsection, it was shown that the magnitude of heterogeneity affects wormhole structure but its inuence on optimum
Damkohler number is not signicant. The dissolution pattern
produced is observed to depend on the relative magnitudes of
convection, reaction and dispersion in the system. Because of
the large variation in injection velocities (over three orders of
magnitude) in core experiments, different transport mechanisms become important at different injection velocities, each
leading to a different dissolution pattern. For example, at high
injection velocities convection is more dominant than dispersion and it leads to uniform dissolution, whereas at low injection velocities dispersion is more dominant than convection
leading to face dissolution. A balance between convection,
reaction and dispersion levels in the system produces worm-
3239
(32)
lT
0.1 to 1 f Wormhole
lX
(33)
lT
O1 f uniform dissolution
lX
(34)
and
C f
keff Cf
x
(35)
where u tip is the velocity of the uid at the tip of the wormhole
and k eff is an effective rate constant dened as
D eT
k eff
(37)
l T k effDeT
lX
utip
(38)
k ck s
aD
k s k c eT
u tip
(39)
u tip
lT
O1 f Face dissolution
lX
2C f
k effCf
y 2
u opt
k ck s
a D k effD eT
k s k c eT
(40)
1
1
1
k eff
k sa k ca
Pe L
Thus, the length scale over which the acid is consumed in the
ow direction is given by
Da
2r
1
Sh
A D T 1/ 2M
o A DT 1/ 2 M
(41)
(42)
where
u tip
lX
keff
3240
(36)
December 2005
M u o/u tip
Vol. 51, No. 12
AIChE Journal
lT 1 DT
L A
Da Da
Pe L
k sa oD m
u o2
(43)
lT
L
AIChE Journal
D eT
k effL2
r
Sh
2
DT
A
1/ 2
(44)
December 2005
1/ 2
(45)
From Eq. 45, it can be seen that the front thickness or the
wormhole diameter is inversely proportional to the square root
of macroscopic Thiele modulus 2. Thus, for increasing values
of macroscopic Thiele modulus (or decreasing levels of dispersion), the diameter of the wormhole decreases, thereby
decreasing the minimum pore volume required to breakthrough. Figure 9 shows pore volume of acid required for
breakthrough versus reciprocal of the parameter o for three
different values of 2 for a kinetically controlled reaction
(2 0.07). The minimum pore volume required to breakthrough decreases with increasing values of macroscopic
Thiele modulus 2. Figure 10 shows the nal porosity proles
at the optimum injection rate in Figure 9 for different values of
macroscopic Thiele modulus. It can be seen from the gure that
the wormhole diameter decreases with increasing values of 2.
The earlier analysis shows that optimum injection rate and
minimum pore volume required for breakthrough are determined by o and macroscopic Thiele modulus 2. The breakthrough curves in Figure 9 are plotted again with respect to
Damkohler number Da in Figure 11 for different values of
macroscopic Thiele modulus 2. It can be seen from the gure
that the optimum Damkohler number is dependent on the value
of 2. Thus, changing the value of 2 changes the optimum
Damkohler number whereas the parameter is always of order
unity for different values of 2 (Figure 9). We believe that
gives a better criterion than the optimum Damkohler number
for predicting wormhole formation. As shown in Figure 11, 2
does not affect the number of pore volumes required to breakthrough in the high injection rate regime. This is because
dispersion effects are negligible at high injection rates, where
convection and reaction are the dominant mechanisms. The
slope of the breakthrough curve at low injection rates and the
minimum pore volume are dependent on the value of 2
showing that dispersion becomes an important mechanism at
lower injection rates where wormholing, conical and face dissolution occur. The breakthrough curve for 2 104 shows a
minimum pore volume that is higher than that required for
3241
Cs
Cf
ks
1
kc
Cf
2r
1
Sh
3242
2 [r o 1 m20 m]
2 /Sh
C s /C f
5.3 105
0.00170.034
0.00060.0113
0.990.98
4 106
4.8 105
0.00240.048
0.00080.016
0.990.98
0.00460.092
0.00150.0306
0.990.97
0.010.2
1.1122.2
0.0030.06
0.377.4
0.990.94
0.730.135
D m (cm2/s)
k s (cm/s)
6 106
6 10
4.5 106
3.6 105
1.4 10
2.3 104
2 101
December 2005
AIChE Journal
om P T/Pe L
a oD m2
2u o2r o
(49)
From Eq. 44, it can be shown that the minimum pore volume
depends on the parameter m . Equation 48 shows that structure property relations have a stronger inuence on the optimal
criterion for mass transfer controlled reactions when compared
to kinetically controlled reactions where
PT ShA
c f uc f c f
cf
t
x
y
r
x
x ur
os PT
m2
os PT
cf
y ur
2
y
m
cf
x
y
Da
A D 1/ 2M
Pe L T
where
PT
a oLD m
2u or o
(46)
a oL 2
2r o
(47)
P T ShA D T
Pe L
r
1/ 2
M om
ShA D T
r
1/ 2
(48)
The x-coordinate is independent of reaction rate (parameters: N ac 0.1, o 0.2, f [0.15, 0.15]).
where
AIChE Journal
December 2005
3243
c f
DaA cf
c f
2
x
y
x
r
1
Sh
os Da
cf
x ur
2
os Da
cf
y ur
2
(50)
Figure 14. Pore volume required for breakthrough is inversely proportional to the acid capacity
number (parameters: 2 0.07, o 0.2, f
[0.15, 0.15], 103).
and
3244
December 2005
DaA c f
2r
1
Sh
3
o
K
exp b
Ko
o
1
(51)
(52)
The relations for average pore radius and interfacial area are
given by Eqs. 8 and 9. By changing the value of b in Eq. 52,
the increase in local permeability with porosity can be made
gradual or steep. Figures 15 and 16 show the effect of b on
evolution of permeability and interfacial area with porosity. It
can be seen from the gures that for low values of b, the
changes in permeability and interfacial area with porosity are
gradual until the value of local porosity is close to unity, and
the change is very steep for large values of b. Figure 17 shows
the effect of structure property relations on the breakthrough
curve for very low and large values of b. A mass transfer
controlled reaction is considered in these simulations because
the effect of structure property relations on optimal conditions
Vol. 51, No. 12
AIChE Journal
December 2005
reaction front. This leads to faster development and propagation of wormholes leading to shorter breakthrough times or
lower pore volumes to breakthrough.
Experimental Comparison
Wormholing in carbonates is a 3-D phenomenon. Comparison of model predictions with experimental data requires a
3-D model. Since the current model is 2-D, and due to a lack
of 2-D experimental data in carbonates, the model predictions
are compared to 2-D experiments on salt-packs reported in
Goler et al.26 In those experiments, undersaturated salt solution was injected into solid salt packed in a Hele-Shaw cell of
dimensions 25 cm in length, 5 cm in width, and 1 mm in
thickness. Because the thickness of the cell is very small
compared to the width and the length of the cell, this conguration may be considered 2-D (provided the wormhole diameter is bigger than the thickness, which is the case in these
experiments). The average values of permeability and porosity
of the salt-packs used in the experiments are reported to be
1.5 1011 m2 and 0.36, respectively. Solid salt dissolves in
the undersaturated salt solution, and creates dissolution patterns that are very similar to patterns observed in carbonates.
The dissolution of salt is assumed to be a mass transfer controlled process (Remark: This assumption breaks down as the
salt concentration in the injected solution approaches saturation
value). Figure 18 shows the experimental data on pore volumes
of salt solution required to breakthrough at different injection
rates for two different inlet concentrations (150 g/L and 230
g/L) of salt solution. The saturation concentration (C sat ) of salt
is 360 g/L and the density of salt ( salt ) is 2.16 g/cm3. The
dissolution of salt in an undersaturated salt solution is a process
very similar to dissolution of carbonate due to reaction with
Vol. 51, No. 12
3245
acid, and the model developed here can be used for salt
dissolution by dening the acid concentration to be C f
C sat C salt . Thus, the acid capacity number for a salt solution
of concentration C o g/L is given by
N ac
C sat C o
salt
Using the above equation, the acid capacity numbers for salt
concentrations of 230 g/L and 150 g/L are calculated to be 0.06
and 0.097 respectively.
To compare model predictions with experimental data, we
need information on initial average pore radius, interfacial area,
and structure-property relations. This data is difcult to obtain
directly and, hence, the model is calibrated with experimental
data to obtain these parameters. Using these parameters, the
model is simulated for a different set of experimental data for
comparison. From the earlier results, we know that for masstransfer controlled reactions, the pore volumes of salt solution
required to breakthrough is a function of the parameters om ,
2
m
, and structure property relations for a given inlet concentration. The model is rst calibrated to the breakthrough curve
corresponding to the inlet salt solution concentration of 150
g/L. For calibration, the largest uncertainty arises from lack of
information on structure-property relations, so the relation in
Eq. 52 is used with the value of b 1. The minimum pore
2
volume to breakthrough depends on m
, and its value is used
to calibrate to the experimental minimum after the structure
property relations are xed. Then, the pore volume to breakthrough curve is generated for different values of om . Figure
19 shows the calibration curve of the model with the experimental data. The value of m used for calibration is 534. This
value of m is used to simulate the model for inlet salt
concentration of 230 g/L (N ac 0.06). The comparison of
3246
December 2005
AIChE Journal
December 2005
Acknowledgments
This work was partially supported by an unrestricted grant by Schlumberger to the University of Houston. We thank Drs. Eduard Siebrits and
Philippe Tardy for many helpful comments.
Literature Cited
1. William B, Gidley J, Schechter RS. Acidizing Fundamentals. SPE
Monograph Series; 1992.
2. Schechter RS. Oil Well Stimulation. New Jersey: Prentice Hall Inc.;
1992.
3. Economides M, Hill A, Ehlig-Economides C. Petroleum Production
Systems. Prentice Hall Inc.; 1993.
4. Economides M., Nolte K (Eds.). Reservoir Stimulation. Wiley; 2000.
5. Daccord G, Lenormand R. Fractal patterns from chemical dissolution.
Nature. 1987; 325: 41 43.
6. Daccord G. Chemical dissolution of a porous medium by a reactive
uid. Phys Rev Lett. 1987; 58: 479 482.
7. Daccord G, Lenormand R, Lietard O. Chemical dissolution of a porous
medium by a reactive uid-1: model for the wormholing. Chem Eng
Sci. 1993a; 48: 169 178.
8. Daccord G, Lenormand R, Lietard O. Chemical dissolution of a porous
medium by a reactive uid-2: convection vs reaction, behavior diagram.
Chem Eng Sci. 1993b; 48: 179 188.
9. Hoefner ML, Fogler HS. Pore evolution and channel formation during
ow and reaction in porous media. AIChE J. 1988; 34: 4554.
10. Wang Y, Hill AD, Schechter RS. The optimum injection rate for
matrix acidizing of carbonate formations. SPE 26578. SPE Ann. Tech.
Conf. and Exhib., Houston, TX; 1993; 675 687.
11. Frick TP, Mostozadeh B, Economides MJ. Analysis of radial core
experiments of hydrochloric acid interaction with limestones. SPE
27402. SPE Intl. Symposium on Formation Damage Control, Lafayette,
LA; 1994;577592.
12. Fredd CN, Fogler HS. Inuence of transport and reaction on wormhole
formation in carbonate porous media. AIChE J. 1998; 44: 19331949.
13. Bazin B. From matrix acidizing to acid fracturing: A laboratory
evaluation of acid/rock interactions. SPE Prod. & Facilities. Feb. 2001;
2229.
14. Buijse MA. Understanding wormholing mechanisms can improve acid
treatments in carbonate formations. SPE Prod. & Facilities. Aug. 2000;
168.
15. Fredd CN, Fogler HS. Optimum conditions for wormhole formation in
carbonate porous media: Inuence of transport and reaction. SPE J.
1999; 4: 196.
16. Fredd CN. Dynamic model of wormhole formation demonstrates conditions for effective skin reduction during carbonate matrix acidizing.
SPE 59537. SPE Permian Basin Oil and Gas Recovery Conference.
Midland, Texas; 2000: 1.
17. Daccord G, Touboul E, Lenormand R. Carbonate Acidizing: Toward a
quantitative model of the wormholing phenomenon. SPE Prod Eng.
1989; 63.
18. Hung KM, Hill AD, Sepehrnoori K. A mechanistic model of wormhole growth in carbonate matrix acidizing and acid fracturing. J Pet
Tech. 1989; 41: 59.
19. Huang T, Hill AD, Schechter RS. Reaction rate and uid loss: The
keys to wormhole initiation and propagation in carbonate acidizing. SPE
37312. SPE Intl. Symposium on Oileld Chemistry, Houston, Texas
1997; 775.
20. Huang T, Zhu D, Hill AD. Prediction of wormhole population density
in carbonate matrix acidizing. SPE 54723. SPE European Formation
Damage Conference; 1999;1 The Hague, Netherlands.
21. Gdanski R. A fundamentally new model of acid wormholing in carbonates. SPE 54719. SPE European Formation Damage Conference;
1999: 1 The Hague, Netherlands.
22. Panga MKR, Ziauddin M, Gandikota R, Balakotaiah V. A new model
for predicting wormhole structure and formation in acid stimulation of
carbonates. SPE 86517. SPE International Symposium and Exhibition
on Formation Damage Control; 2004;1 Lafayette, Louisiana.
23. Liu, X, Ormond A, Bartko K, Li Y, Ortoleva P. A geochemical
3247
3248
December 2005
Appendix A
Numerical algorithm
The equations are discretized on a 2-D domain using a
control volume approach. While discretizing the species balance equation, an upwind scheme is used for the convective terms
in the equation. The following algorithm is used to simulate ow
and reaction in the medium. The pressure, concentration and
porosity proles in the domain at time t are denoted by pt, ct and
t. Porosity and concentration proles in the domain are obtained
for time t t (ctt, tt), by integrating the species balance
and porosity evolution equations simultaneously using the ow
eld calculated from the pressure prole ( pt) by applying Darcys
law. Integration of concentration and porosity proles is performed using Gears method for initial value problems. The calculation for concentration and porosity proles is then repeated for
time thalf t t/2 using the velocity prole at time t. The ow
eld at t t/2 is then calculated using the concentration prole
chalf and porosity prole half. Using the ow prole at thalf the
values of concentration and porosity are again calculated for time
t t, and are denoted by cnew and new. To ensure convergence,
the norms ctt cnew and tt new are maintained below
a set tolerance. If the tolerance criterion is not satised the calculations are repeated for a smaller time step. The earlier procedure
is repeated until the breakthrough of the acid which is dened as
the decrease in the initial pressure by a factor 100.
Manuscript received Feb. 16, 2005, and revision received Apr. 25, 2005.
AIChE Journal
Hota V. S. GangaRao
Dept. of Civil and Environmental Engineering and Constructed Facilities Center, West Virginia University,
Morgantown, WV 26506
L. N. Sridhar
Dept. of Chemical Engineering, University of Puerto Rico, Mayaguez, PR 00681
DOI 10.1002/aic.10605
Published online August 22, 2005 in Wiley InterScience (www.interscience.wiley.com).
Keywords: diffusivity measurement, moisture permeation, vinyl ester, polymer nanocomposites, montmorillonite, barrier properties
Introduction
Glass-berreinforced polymer (GFRP) composites incorporating thermosetting polymer matrices are now being used in
the construction and repair of bridges and other civil structures.1 However, widespread use of GFRPs in construction,
where a service life in excess of 50 years is desired, has been
hindered by the lack of long-term durability and performance
Correspondence concerning this article should be addressed to R. K. Gupta at
rakesh.gupta@mail.wvu.edu.
AIChE Journal
December 2005
data on which to base design calculations. A particular difculty has been that atmospheric moisture can diffuse to the
bermatrix interface and cause both delamination and ber
weakening.2 One way to reduce moisture ingress is to reduce
the effective moisture diffusion coefcient, and this can be
achieved by dispersing clay platelets in the matrix polymer,
forming a nanocomposite; the diffusing molecules have to
travel around the akelike barriers, and the additional path
length, by itself, is predicted to substantially decrease the
diffusivity (see, for example, Cussler et al.3 and Yano et al.4).
Note that the most common variety of clay for nanocomposite
synthesis is montmorillonite.5,6 It is a swelling clay consisting
Vol. 51, No. 12
3249
December 2005
Material Characterization
As reported previously,5 X-ray diffraction scans on the nanocomposite samples were indicative of either exfoliation or
breakdown of the ordered structure of clay. Low-magnication
transmission electron micrographs of the same samples showed
that the clay was reasonably uniformly distributed, and the clay
platelets were randomly oriented. High-magnication micrographs, however, revealed the presence of aggregates made up
of two to ten platelets; the distance between adjacent platelets
in these aggregates ranged from 4.4 to 5 nm, whereas the
platelet length ranged from 80 to 180 nm. Thus, the clay is only
partially exfoliated, and this has consequences for the resulting
barrier properties.
AIChE Journal
Figure 1. (a) SEM of as-prepared GFRP sample; (b) SEM of GFRP sample after immersion in distilled water at room
temperature for 2 months; (c) SEM of GFRP sample containing 5 wt % Cloisite 10A after immersion in
distilled water at room temperature for 2 months.
ity from either sorption experiments or permeation experiments. In the present work, transient diffusion tests were
conducted by immersing the various samples in distilled
water or alkali of pH 13 or 0.1 M salt solution at a constant
temperature ranging from 4 to 62C. Beakers containing
these solutions were closed at the top with aluminum foil to
prevent evaporation. A refrigerator was used to conduct
experiments at temperatures below room temperature,
whereas an oven was used at temperatures above room
temperature. The samples were periodically removed, blotAIChE Journal
December 2005
3251
2c
c
D 2
t
x
(1)
Mt
1
M
8
D(2n 1)2 2 t
exp
(2n 1) 2 2
4l2
(2)
Dt
Mt
4
M
(2l) 2
1/ 2
(3)
December 2005
c
nx
1
t
x
(4)
t
1 x 2
(5)
AIChE Journal
D1
1
(6)
D s/D t 1
(7)
Dt
Finally, then
December 2005
3253
Equilibrium
Moisture
Content (wt %)
System
4C
25C
4C
25C
Neat resin
1 wt % Cloisite 10A
2 wt % Cloisite 10A
5 wt % Cloisite 10A
0.14
0.11
0.11
0.07
0.74
0.40
0.28
0.06
0.60
0.85
0.90
0.97
0.66
1.13
1.34
2.08
December 2005
Equilibrium
Moisture
Content (wt %)
System
4C
28C
4C
28C
Neat resin
1 wt % Cloisite 10A
2 wt % Cloisite 10A
5 wt % Cloisite 10A
0.12
0.12
0.08
0.08
0.66
0.43
0.37
0.26
0.58
0.66
0.59
0.61
0.68
0.88
0.95
1.09
AIChE Journal
December 2005
(8)
Neat resin
1 wt % Cloisite 10A
2 wt % Cloisite 10A
5 wt % Cloisite 10A
0.75
0.70
0.67
0.64
3255
5.
6.
7.
Conclusions
The results of this research demonstrate that montmorillonite
can reduce moisture transport through polymers such as vinyl
ester. The measured reduction in moisture diffusivity on clay
addition, though, appears to be larger in sorption experiments
than that in permeation experiments, and this seems to be
related to the fact that results of sorption experiments are not
very sensitive to the extent of ller exfoliation when the diffusing molecule is strongly adsorbed on the ller surface.
Permeation experiments are thus preferable to sorption experiments, and these might even be used to estimate the effectiveness of a given processing technique to exfoliate clay platelets.
8.
9.
10.
11.
12.
13.
Acknowledgments
This research was funded by the U.S. Department of Transportation and
the Federal Highway Administration under contract number DTFH61-C00021. Professor Sridhar was on sabbatical leave at West Virginia University when this research was done.
14.
15.
Literature Cited
1. Freed AD, Gupta RK, GangaRao HVS. Minimizing voids in pultruded
polymer composites. Proc of the Society of Plastics Engineers, ANTEC, May 4 8, Nashville, TN; 2003:2105-2109.
2. Kajorncheappunngam S, Gupta RK, GangaRao HVS. Effect of aging
environment on degradation of glass-reinforced epoxy. ASCE J Compos Construct. 2002;6:61-69.
3. Cussler EL, Hughes SE, Ward WJ, Aris R. Barrier membranes. J
Membr Sci. 1988;38:161-174.
4. Yano K, Usuki A, Okada A, Kurauchi T, Kamigaito O. Synthesis and
3256
December 2005
16.
17.
Manuscript received Apr. 14, 2004, and revision received Apr. 4, 2005.
AIChE Journal
Introduction
Gradients of temperature or molar fraction (more general:
electrochemical potential) along the boundary of a liquid
phase, or between this boundary and the bulk, can give rise to
Present addresses: S. Post, BASF Aktiengesellschaft, GIC/PQ920, 67056 Ludwigshafen, Germany (stephan.post@basf-ag.de); I. Urukova, Applied Thermodynamics, University of Kaiserslautern, Erwin-Schrodinger-Strae, D-67663 Kaiserslautern,
Germany (urukova@mv.uni-kl.de).
Correspondence concerning this article should be addressed to E. Tsotsas at
evangelos.tsotsas@vst.uni-magdeburg.de.
AIChE Journal
December 2005
3257
December 2005
AIChE Journal
Figure 1. Apparatus for the investigation of the evaporation of a binary mixture from planar gaps into
inert gas.
December 2005
3259
Figure 3. Apparatus for the investigation of the evaporation of a binary mixture from packed beds
into inert gas.
December 2005
(1)
AIChE Journal
d p2 3
1
150 1 2
(2)
December 2005
3261
l,0 D l/G D
(3)
December 2005
AIChE Journal
transfer coefcients g,i are determined with the help of pureliquid evaporation experiments.
Contrary to the treatment of the gas phase, Stefan uxes
must not be neglected in the liquid, so that the relationship
N i n lA Ph lr iln
ri xi,Ph
ri xi,B
Ni
ri
N
(8)
Ki
y i,Ph i x i,Ph p i0
x i,Ph
p
(9)
model for mass and heat transfer in the gas and the liquid phase
has been used. The respective mass balances for the liquid are
N1 N2 N
Ni
dN1
dt
d
N x
dt l i,B
(4)
(5)
with i 1, 2 the component index and N i the molar evaporation rates. For the gas it is
N i N gY i,out
(6)
with N g the molar ow rate of the inert carrier (air), and Y i,out
the molar content of vapors at the outlet of this stream (Y i,in
0).
Because of the already discussed high dilution, gas-side
mass transfer kinetics may be linearized and expressed in terms
of molar content; (for y i 1 it is: Y i y i ). For the ideally
backmixed gas phase it, then, holds
N i n gA Ph g,iY i,Ph Y i,out
(7)
Herein APh is the gasliquid phase contact area (crosssectional area of the gaps). It has been pointed out that the mass
AIChE Journal
December 2005
3263
N h T
2
N gh g,in h g,out
i g,i
Ph
g0
Q
(10)
i1
and
N h T
2
l
Q
v,i
g
c p,l,iT Ph T l,B Q
Ph
(11)
i1
g ) and
respectively. The heat ow rates coming from the gas (Q
(12)
(13)
The dimensionless quantities KA,g and KA,l, so-called Ackermann corrections that account for the temperature change of
the evaporation streams in the boundary layers, are for the
conditions of the present investigation approximately equal to
unity.41
The gas-side heat transfer coefcient g is calculated according to the Lewis analogy
g
n gc p,gLein
g,i
(14)
g
ng cp,g Dg,i
(15)
with
Lei
from the gas-side mass transfer coefcients g,i and the exponent n, as determined separately from pure-liquid evaporation
experiments. In Eq. 15 Dg,i are the binary diffusion coefcients
of isopropanol (index: 1) and water (index: 2) in air, g and c p,g
are the thermal conductivity and the molar specic capacity of
the gas mixture. In this context, notice that material properties
referring to the phase bulks (index: B) or to the gasliquid
interface (index: Ph) are always calculated at the local values
of concentration and temperature, whereas material properties
referring to the boundary layers are calculated at the respective
arithmetic averages.
The liquid-side heat transfer coefcient l has been related
to the liquid-side mass transfer coefcient l according to
l
3264
l l
Dl
(16)
December 2005
Eh
l
l,0
(17)
AIChE Journal
the change of selectivity arising from deviations from isothermal conditions is moderate (Figure 8). Consequently, inaccuracies in the calculation of the liquid-side heat transfer coefcient l according to Eq. 16 are not expected to be crucial for
the described determination of enhancement factors. This sit-
December 2005
Figure 11. Adiabatic temperatures (l 3 0) of the mixture isopropanolwater in air of various bulk
temperatures for l 3 0 and l 3 .
Vol. 51, No. 12
3265
Ra
l gGD3
l Dl
(19)
l GD
l Dl
(20)
Ma
(18)
Figure 13. Separate presentation of Rayleigh and Marangoni numbers from Figure 12.
Figure 12. Enhancement factors vs. the sum of the Rayleigh and Marangoni number for the experimental data of Figure 4, that is, at different
temperatures in the bulk of the gas phase for
x1,B,0 0.70.
l,0
3266
l
GD
December 2005
AIChE Journal
Figure 14. Measured and calculated enhancement factors for interfacial convection in planar gaps.
In the next step, the equivalent Reynolds number is considered to be proportional to the enhancement factor, the Schmidt
number is replaced by a more general damping function (DSc),
and factor and exponent of Eq. 22 are given free for tting.
Several trials41 have led to the damping function
D Sc
10538.55
10543.43
x 0.08358
1 exp
0.067567
with
1
w 2 g lG D l/G D
2 l
(21)
x
Ra Ma
Re 2
Sc
2
eq
(22)
Eh 5 10 3
(23)
l
Dl
(24)
This is analogous to the usual derivation of simplied correlations for heat or mass transfer by combined forced and free
convection, shows that the use of the sum Ra Ma is not
completely arbitrary, and introduces Sc as the additional inuence parameter. On the other hand, it can only be coarsely
approximate. Indeed, even in the case of conventional forced
and free convection, nonlinear, supporting, or competitive
combinations are well known.45
AIChE Journal
(26)
(25)
December 2005
Ra Ma
DSc
4/5
(27)
3267
tion is possible, arising from preferential evaporation of alcohol, and may assist the thermally initiated interfacial convection. Gas bulk temperature loses its prevailing role as the
operating parameter controlling selectivity and convection.
This is similar to the observations with planar gaps (compare
with Figure 5). A further similarity exists in the inhibition of
the mixing inuence of micro convective ow by lateral space
restriction. The latter has been realized in the case of packed
beds by systematic decrease of the particle diameter dp. As the
data of Figure 17 show, the evaporation was nonselective for
the smallest particles with dp 0.23 mm, which is equivalent
to the complete absence of interfacial convection. An inhibition
of the mixing inuence of interfacial convection has also been
observed by increasing the bed height h, as the data of Figure
18 exemplify. This trend was valid for any combination of the
remaining operating and geometrical parameters. In this respect, a difference of behavior appears to exist between packed
beds and planar gaps because the impact of gap depth on
interfacial convection has been found to be ambiguous in the
case of planar gaps (see Figure 7).
Although the further evaluation of measured data was in
principle the same as that for planar gaps, some modications
have been carried out in the kinetic equations of the model, and
in the denition and calculation of dimensionless numbers.
These modications will be discussed in the following.
Concerning the model, not the cross-sectional area of the
gaps, but the total cross-sectional area of the space lled with
the packed bed (A0) is used in the kinetic equations for gas-side
mass transfer, liquid-side mass transfer, gas-side heat transfer,
and liquid-side heat transfer leading, after omission of the
Ackermann factors, to
N i n g A 0 g,iY i,Ph Y i,out
N i n lA 0 lr iln
ri xi,Ph
ri xi,B
(28)
(29)
December 2005
Figure 16. Concentration curves for the same conditions as in Figure 15, although at x1,B,0 0.30.
Vol. 51, No. 12
AIChE Journal
D l,bed D l1 1
(34)
l
Figure 17. Concentration curves for the evaporation of a
binary mixture from packed beds into inert
gas at different particle diameters (dp); measurements, limiting cases, and calculations
(solid lines) by punctual adaptation of the liquid-side mass transfer coefcient l.
g gA 0T g,B T Ph
Q
(30)
l lA 0T l,B T Ph
Q
(31)
l l
D l,bed
(35)
/1
(32)
l,0
D l,bed
h
(33)
with
AIChE Journal
December 2005
3269
Mas
Figure 19. Enhancement factors vs. the sum of the Rayleigh and Marangoni number for the experimental data of Figure 17 (different particle
diameters, dp).
beds than for the planar gaps. Thus, the approximate character
of Eq. 35 will not bias the evaluation of the disturbed state of
the system.
More critical is the consideration, or not, of liquid-side heat
transfer for the undisturbed state of the system. Respective
calculations have been carried out at the adiabatic limit of l 3
0 (as for the planar gaps), but also with heat transfer coefcients after
l,0
bed
h
(36)
l gh
l Dl,bed
(37)
December 2005
l
l Dl,bed h
(38)
Figure 20. Separate presentation of Rayleigh and Marangoni numbers from Figure 19.
Vol. 51, No. 12
AIChE Journal
Table 1. Marangoni and Rayleigh Numbers for All Packed-Bed Experiments of the Present Investigation without Detectable
Interfacial Convection (Nonselective Evaporation, Eh 1)
Varied Operating Parameters
x 1B,0
T g,B (K)
d p (mm)
h (mm)
Mas
Ras
Mas Ras
Eh
0.7
0.7
0.7
0.7
0.7
343.15
303.15
303.15
303.15
303.15
3.5
0.23
0.23
0.23
0.57
15.0
5.0
15.0
25.0
25.0
negative
8.76 100
2.68 100
1.69 100
1.04 101
negative
5.54 101
1.62 102
2.78 102
1.71 103
negative
6.42 101
1.65 102
2.79 102
1.72 103
1.0
1.0
1.0
1.0
1.0
(39)
2
d
31 p
Reeq
udp
1 l
(43)
(44)
In the next step, the equivalent Reynolds number is considered to be proportional to the enhancement factor, and the
factor and exponent of Eq. 41 are tted, yielding
Eh 29.47
Ras Mas
Sc
4/5
(45)
(40)
2
Reeq
166.67
Ras Mas
Sc
(41)
is obtained, with
31 h
2dp
(42)
Figure 21. Measured and calculated enhancement factors for interfacial convection in packed
beds.
December 2005
3271
December 2005
rangoni convection below the azeotrope. Convection is inhibited by lateral space restriction and was thus found to decrease
with decreasing gap width or particle diameter.
Derived enhancement factors are much larger than previously reported in the literature (up to almost 100 for the sliced
slabs and to almost 2800 for the packed beds, compared to
previously documented values of maximally about ve). Although our presented correlation attempts are still coarse and
preliminary, they do indicate a strong dependency of the enhancement factor on the Rayleigh or Marangoni number, with
an exponent of 0.8 instead of, typically, 0.5 in the literature.
This is explained by the fact that the experiments were conducted in a region of high intensication of liquid-side mass
transfer that had never been previously investigated.
The appropriate consideration of structure and permeability
of the porous medium has been discussed, along with the
denition of DArcy-modied Rayleigh and Marangoni numbers for the packed beds. It has been pointed out that the
present evaluation is, with respect to liquid-side heat transfer,
close to the isothermal limit for the disturbed and adiabatic for
the undisturbed state of the system. Although the former is
reasonable and not very important for the result of the evaluation, the latter may be problematic; however, this is hardly
avoidable in the frame of a simplied one-dimensional model.
Finally, four nonselective experiments with packed beds have
been identied, which correspond to small, but positive values
of the dimensionless driving potentials, underlining the stability character of the problem under consideration.
One obvious continuation of the present work would be to
systematically identify critical Rayleigh and Marangoni numbers by use of such methods as smaller increments in the
variation of gas bulk temperature above the azeotropic point.
Such experiments must be combined with a rigorous stability
analysis andin the perspectivewith a more fundamental,
two-dimensional modeling of transport phenomena in the porous obstacle. On the other hand, the developed experimental
method allows for a systematic variation of liquid properties,
either by additives or by use of alternative binary mixtures. In
this way, a better understanding of the underlying, complex
phenomena may be achieved, along with better, more accurate
correlations.
In total, the presented method may be seen as complementary to previous experimental approaches referring closely to
processes such as liquidliquid extraction, absorption or distillation complementary in the sense of enabling the investigation of otherwise inaccessible regions of high enhancement of
liquid-side mass transfer. Consequently, and by extrapolation,
it may contribute to a better understanding of the inuence of
interfacial phenomena on such processes. However, the used
porous obstacles are not of interest only as articial models for
thick laminar boundary layers, but also for their own sake as
membranes, catalysts, products to be dried, and soils in numerous chemical and environmental engineering applications.
Acknowledgments
The described research was supported in part by the postgraduate program of the German federal state of SaxonyAnhalt and by the MaxBuchner-Foundation.
Notation
Aph gasliquid phase contact area, m2
AIChE Journal
A0
c p
d
D
DSc
Eh
g
G
h
h
h v
K
KA
Le
n
N
N
Ma
Mas
p
Q
r
Ra
Ras
Reeq
Sc
t
T
u
V
w
x
x
y
Y
Literature Cited
1. Kazhdan D, Shtilman L, Golovin AA, Prismen LM. Non-linear waves
and turbulence in Marangoni convection. Phys Fluids. 1995;7:26792685.
2. Severin J, Herwig H. Onset of convection in the RayleighBenard ow
with temperature dependent viscosity: An asymptotic approach. Z
Angew Math Phys. 1999;50:375-386.
3. Bragard J, Velarde MG. BenardMarangoni convection: Plan forms
and related theoretical predictions. J Fluid Mech. 1998;368:165-194.
4. Birikh RV, Briskman VA, Rudakov RN, Velarde MG. Marangoni
Benard instability of a oating liquid layer with an internal, permeable,
heated or cooled divider and two deformable, open surfaces. Int J Heat
Mass Transfer. 1995;38:2723-2731.
5. Or AC, Kelly RE. Onset of Marangoni convection in a layer of uid
modulated by a weak nonplanar oscillatory shear. Int J Heat Mass
Transfer. 1995;38:2269-2279.
6. Golovin AA, Nepomnyashchy AA, Pismen LM, Riecke H. Steady and
oscillatory side-band instabilities in Marangoni convection with deformable interface. Physica D. 1997;106:131-147.
7. Hashim I, Wilson SK. The onset of oscillatory Marangoni convection
in a semi-innitely deep layer of liquid. Z Angew Math Phys. 1999;
50:546-558.
8. Boeck T, Thess A. BenardMarangoni convection at low Prandtl
number. J Fluid Mech. 1999;399:251-275.
9. Li M, Zheng D, Zhu T. Instability of the Marangoni convection in a
liquid bridge with liquid encapsulation under microgravity condition.
Int J Heat Mass Transfer. 2002;45:157-164.
10. Tan KK, Thorpe RB. The onset of convection driven by buoyancy
effects caused by various modes of transient heat conduction: Part I.
Transient Rayleigh numbers; Part II. The sizes of plumes. Chem Eng
Sci. 1999;54:225-238,239-244.
11. Gottberg von FK, Hatton TA, Smith KA. Surface instabilities due to
interfacial chemical reaction. Ind Eng Chem Res. 1995;34:3368-3379.
12. Mendes-Tatsis MA, Perez de Ortiz ES. Marangoni instabilities in
systems with an interfacial chemical reaction. Chem Eng Sci. 1996;
51:3755-3761.
13. Kang KH, Choi CK, Hwang IG. Onset of solutal Marangoni convection in a suddenly desorbing liquid layer. AIChE J. 2000;46:15-23.
14. Skarda JRL, McCaughan FE. Exact solution to stationary onset of
convection due to surface tension variation in a multicomponent uid
layer with interfacial deformation. Int J Heat Mass Transfer. 1999;42:
2387-2398.
15. Bau HH. Control of MarangoniBenard convection. Int J Heat Mass
Transfer. 1999;42:1327-1341.
16. Sternling CV, Scriven LE. Interfacial turbulence: Hydrodynamic instability and the Marangoni effect. AIChE J. 1959;5:514-523.
17. Perez de Ortiz ES, Sawistowski H. Interfacial stability of binary
liquidliquid systems, Part I: Stability analysis. Chem Eng Sci. 1973;
28:2051-2062.
18. Imaishi N, Fujinawa K. Theoretical study of the stability of two-uid
layers. J Chem Eng Jpn. 1974;7:81-86.
19. Reichenbach J, Linde H. Linear perturbation analysis of surfacetension-driven convection at a plane interface (Marangoni instability).
J Colloid Interface Sci. 1981;84:433-438.
20. Wassmuth F, Laidlaw WG, Coombe DA. Interfacial instabilities: The
Linde instability. Chem Eng Sci. 1990;45:3483-3490.
21. Wolf S. Phasengrenzkonvektionen beim Stoffubergang in FlussigFlussig-Systemen (PhD Thesis, Technical Univ. of Munich).
Fortschritt-Berichte VDI, Series 3, No. 584. Dusseldorf, Germany:
VDI-Verlag; 1999.
22. Chai AT, Zhang N. Experimental study of MarangoniBenard convection in a liquid layer induced by evaporation. Exp Heat Transfer.
1998;11:187-205.
23. Guzun-Stoica A, Kurzeluk M, Floarea O. Experimental study of Marangoni effect in a liquidliquid system. Chem Eng Sci. 2000;55:38133816.
24. Pertler M, Haberl M, Rommel W, Bla E. Mass-transfer across liquidphase boundaries. Chem Eng Process. 1995;34:269-277.
25. Henschke M, Pfennig A. Mass-transfer enhancement in single-drop
extraction experiments. AIChE J. 1999;45:2079-2086.
26. Semkov K, Kolev N. On the evaluation of the interfacial turbulence
(the Marangoni effect) in gas (vapour)liquid mass transfer: Part I. A
method for estimating the interfacial turbulence effect; Part II. Mod-
Greek letters
Indices
ad
bed
B
D
g
i
in
l
out
p
pore
Ph
W
0
0
0
0
adiabatic
bed
bulk
depth
gas
species (1: isopropanol, 2: water)
inlet
liquid
outlet
particle
pore
interface
width
initial
saturation
undisturbed (stable) state
standard conditions
AIChE Journal
December 2005
3273
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
3274
December 2005
38. Sun ZF, Yu KT, Wang, SY, Miao YE. Absorption and desorption of
carbon dioxide into and from organic solvents: Effects of Rayleigh and
Marangoni instability. Ind Eng Chem Res. 2002;41:1905-1913.
39. Riede T, Schlunder EU. Selective evaporation of a binary mixture into
dry or humidied air. Chem Eng Process. 1990;27:83-93.
40. Schwarzbach J, Nilles M, Schlunder EU. Microconvection in porous
media during pervaporation of a liquid mixtureAn experimental
study. Chem Eng Process. 1987;22:163-175.
41. Post S. Zum Einu von Grenzachenkonvektionen auf die Stoffubertragung am Beispiel der selektiven Verdunstung eines binaren Gemisches aus ebenen Spalten (PhD Thesis, Univ. of Magdeburg).
Fortschritt-Berichte VDI, Series 3, No. 722. Dusseldorf, Germany:
VDI-Verlag; 2002.
42. Urukova I. Einu der Mikrokonvektion auf Selektivitat und Stofftransport wahrend der Verdunstung binarer Flussigkeiten aus porosen
Schichten. PhD Thesis. Magdeburg, Germany: Univ. of Magdeburg;
2005.
43. Henneberg M, Ziad Saghir M, Rednikov A, Legros JL. Porous media
and the BenardMarangoni problem. Transport Porous Media. 1997;
27:327-355.
44. Press WH, Flannery BP, Teukolsky SA, Vetterling WT. Numerical
Recipes. Cambridge, UK: Cambridge Univ. Press; 1989.
45. Klan H. Warmeubergang durch Mischkonvektion an umstromten Korpern. VDI-Warmeatlas, Section Fe. 9th Edition. Berlin: SpringerVerlag; 2002.
46. Winterberg M, Tsotsas E, Krischke A, Vortmeyer D. A simple and
coherent set of coefcients for modelling of heat and mass transport
with and without chemical reaction in tubes lled with spheres. Chem
Eng Sci. 2000;55:967-979.
47. Tsotsas E. Heat transfer to gassolid systems: Stagnant packed beds.
Heat Exchangers Design Update. Section 2.8.1. New York, NY:
Begell House; 2000.
Manuscript received May 11, 2004, and revision received Mar. 18, 2005.
AIChE Journal
THERMODYNAMICS
A universal prediction method for Henrys constants of alkanes in alkane solvents has
not been established. The present study proposes a universal method for the prediction of
Henrys constants representing all of the data for alkane/alkane binaries in the literature.
The method involves two universal constants and works at Tri 3, where Tri denotes the
reduced temperatures of the solute alkanes. One of the constants modies the expression
of the innite-dilution activity coefcients for their application to simple uids reecting
end-to-sphere molecular orientations, whereas the other represents the vapor pressures of
hypothetical liquids dened above solute critical. The prediction method also satisfactorily predicts the Henrys constants of rare gases and liquids having spherical molecules,
such as Ar, Kr, Xe, and CCl4, in alkane solvents. Using the prediction method, the most
reliable data of high-pressure vaporliquid equilibrium for alkane/alkane binaries under
dilute conditions in the literature were identied. 2005 American Institute of Chemical
Engineers AIChE J, 51: 32753285, 2005
Introduction
Henrys law constants are of practical importance for the rational design of separation devices in the petroleum and petrochemical industries. Henrys law constants of an alkane solute i (Hi )
and the innite-dilution activity coefcients of solute i (i ), dissolved in alkane solvents, are of profound theoretical interest in
thermodynamics because it is an important simplicity that the
dispersion force is the only attractive molecular interaction acting
on the alkane molecules. On the basis of these insights, a great
number of experimental data for Hi in alkane/alkane binaries have
been reported in the literature. However, at present, a method
predicting Hi for the alkane/alkane binaries covering a variety of
solutesolvent combinations has not been established.
Prediction methods for H i using an equation of state (EOS)
seem to have limitations in applicability arising from mixing
rules. From thermodynamics, H i is given as the product of i
S. Katos e-mail address is Kato-satoru@c.metro-u.ac.jp.
AIChE Journal
December 2005
3275
Prausnitz and Shair,4 Yen and McKetta Jr.5 reported that the
relationships between the nondimensional fugacities of the
saturated vapors for pure liquids at 101.3 kPa and the reduced
temperature of solute i (Tri) are independent of the solute types.
They used the regular-solution theory to express ln i and
determined the fugacities from the solubility data of the gases.
However, for the alkane/alkane binaries, the regular-solution
theory fails to predict reliable values of ln i because the
theory provides a relationship, ln i 0, in spite of the
experimental proof: ln i 0.3 The method by Rodriguez and
Patterson2 has limitations in its applicability because a hypothetical liquid has not been dened. Chappelow III and
Prausnitz1 dened hypothetical liquids by extending the fugacity expressions given by Prausnitz and Chueh6 below the solute
critical to temperatures above the solute critical. This method
does not lead us to a universal prediction expression because
the parameters are dependent on solute type. The reason that
none of the conventional studies has established a universal
prediction method for H i of the alkane/alkane binaries is
ascribed to the fact that they could not use an appropriate
expression for ln i .
Kato et al.3 represented ln i data for solute alkanes in
longer carbon-chain alkanes according to a rst-order function
of the inverse absolute temperature. The alkanes determining
this expression have linear-shaped molecules; therefore, the
expression should be modied for simple uids having spherical molecules, such as methane, and providing different orientation patterns from the linear molecules around the parallel
frameworks formed by the solvent molecules. In the literature,
special attention has not been paid to the molecular orientations
of simple uids, and the same expressions of combinatorial
entropies with those for linear molecules have been used.
However, when the solute approaches a simple uid with
decreasing Ni, a modication of ln i using the acentric factor
proposed by Pitzer7 may be promising for the evaluation of
solution structures and molecular interactions around a molecule of a simple uid. Such a modication has not been found
in the literature. As for the denition of hypothetical liquids, it
seems much more denite to specify hypothetical liquids by their
vapor pressures, given the great amount of knowledge on vapor
pressures that has been accumulated. However, no report has been
found specifying the vapor pressures and the fugacity coefcients
of the saturated vapor phases of hypothetical liquids.
One of the purposes of the present study is to develop a
universal prediction method for H i of alkane/alkane binaries
by the modication of ln i given by Kato et al.3 for application to simple uids and by the denition of hypothetical
liquids in terms of vapor pressures and fugacity coefcients.
The second purpose is to show that the universal prediction
method practically serves to clarify solution structures in alkane/alkane binaries at innite dilution, to verify the reliability
of high-pressure vaporliquid equilibrium (VLE) data for alkane/alkane binaries in the literature, and to predict the H i
values of rare gases in an alkane solvent.
Modeling
Modication of innite-dilution activity coefcients for
solute alkanes approaching a simple uid
Consider a system consisting of a solution of a solute alkane
i innitely diluted in a solvent alkane j and the saturated vapor
3276
December 2005
(1)
(2)
x i 30
whereas H (Pa)
is given as follows8
i
H iPa iPa isp is
(3)
where (Pa)
denotes the innite-dilution activity coefcient of
i
solute i relative to the reference pressure Pa, and specifying the
saturated liquid of solute i at T and at the saturated vapor
pressure pis as a standard state of activity; is denotes the
fugacity coefcient of the saturated vapor of solute i at temperature T. Equations 1 to 3 provide the following expression
for H i
H i i isp isexp
vi (P Pa )
RT
(4)
qi qj
qi
(5)
(6)
q i 20.848 0.540N i 2
(7)
q j 20.848 0.540N j 2
(8)
AIChE Journal
estimated as follows. Figure 1a schematically shows the destruction of the parallel frameworks formed in a solvent alkane
by the introduction of a solute molecule having a linear chain.
When the carbon chain of the solute becomes shorter, the
attractive forces acting between the solute and solvent molecules become weak; therefore, the destructed frameworks of
the solvent do not recover in this case. As demonstrated by Eq.
5, the decrease in (ln i )linear (0 at 298.15 K) with decreasing
Ni reects this growing destruction. On the other hand, as
shown by Figure 1b, when the solute approaches a simple uid,
the parallel frameworks of the solvent molecules are retained
without destruction because a small spherical molecule can
locate at the end of the solvent alkanes long carbon-chain
molecules forming an end-to-sphere orientation; the solution
approaches an ideal solution. Simple uids are characterized by
the acentric factor i, proposed by Pitzer.7 In Figure 2, i is
plotted vs. qi. The straight line drawn in Figure 2 stands for
acentric factors, which are proportional to qi, denoted as (i)linear. The line passes through the points for nonane and decane
having linear molecules. It is now expected that a difference
between i and (i)linear represents the end-to-sphere orientations shown in Figure 1b for simple uids; the present study
assumes the effect of the simple uid on ln i as follows
ln i
m
i
lni linear
(i )linear
i linear
Hi
1
ln is h 1
Tri
Pci i
(11)
(9)
decane
q
qdecane i
(10)
where decane and qdecane denote the acentric factor and the
measure of the molecular surface area of decane, respectively.
The constant m in Eq. 9 is determined from the H i data
satisfying T Tci, where Tci denotes the critical temperature of
solute i. It is stressed that i in Eq. 9 is dened by the activity
specifying the standard state of a pure liquid i at T and pis.
December 2005
3277
Meas.*
N data **
N i /N j
T min
T max
Gjaldbaek (1952)10
Lannung and Gjaldbaek (1960)11
Thomsen and Gjaldbaek (1963)12
Ng et al. (1969)13
Lenoir et al. (1971)14
Jadot (1972)15
Hayduk et al. (1972)16
Hayduk and Castaneda (1973)17
Chappelow, III and Prausnitz (1974)1
Fleury and Hayduk (1975)18
King and Al-Najjar (1977)19
Richon and Renon (1980)20
Parcher and Johnson (1980)21
Monfort and Arriaga (1980)22
De Ligny and van Houwelingen (1986)23
Gonzalez et al. (1987)24
2
1
1
2
2
2
1
1
1
1
1
3
2
2
1
1
1
1
1
50
7
11
10
10
44
4
7
8
20
12
1
22
298.15
298.15
298.15
308.15
298.15
298.15
298.15
298.15
300
258.1
303.1
298.15
298.15
278.1
298.15
298.15
298.15
298.15
298.15
473.15
323.15
298.15
318.1
323.15
475
323.15
343.1
323.15
328.15
323.15
298.15
323.15
1
3
2
1
5
43
18
11
3/10
1/6
3/6
(1,2,3)/18, (1,2,3)/20, (1,2,3)/22
(1,2,3,4)/16, (2,3,4)/17
2/10, 3/(6,7,8,9,10), 4/(6,7,8,9,10)
3/(6,7,12,16)
3/(6,16,17), 4/(6,7,8,12,16,17)
(3,4)/16, (1,2,3,4)/20
3/6
3/(12,14,16)
(1,2,3,4)/16, (1,2,3,4)/18
(1,2,3,4,5)/16
(2,3,4)/10, (2,4)/12
1/10
3/(8,9,10,12,13,14,15),
4/(8,9,10,12,13,14,15)
(3,4)/8
(5,8)/16, (2,4,5,6,7,8,10)/20
(3,4,5)/16, (3,4,5)/24
1/(6,7,8,9,10,11,12,13,14,15,16)
298.15
333
303.15
298.15
343.15
473
373.15
298.15
* Measuring method: 1, static equilibrium cell; 2, gasliquid chromatography; 3, dynamic equilibrium cell.
**Number of data.
Carbon numbers, N i for solute and N j for solvent, are bracketed if the same solvent or solute is used.
lation, ln H i,cal
a bNj. The values of (AAD)ln H, dened as
H i,cal
, dened as (AAD)H (100/Ndata) (H i,exp H i,cal
)/
3278
December 2005
AIChE Journal
p is
Zi 1
dP
P
(12)
December 2005
The value of m
The constant m is determined using the H i data covering
temperatures below the solute critical. The values of the exponential term in Eq. 4 ranged from 0.988 exp[vis(P 101.3)/
RT] 1.012; therefore, in the present study, the exponential
term in Eq. 4 is neglected. To examine the effect of [i/
(i)linear]m on H i , in Figure 5, the H i / i pisis data covering
T Tci is plotted vs. pis/Pci, where m was xed at zero. It is
apparent from Figure 5 that H i / i pisis deviates from unity
when the carbon number decreases. The average absolute deviation (AAD)1, dened by the following equation
AAD1 100/Ndata
H / p
i is
is exp
(13)
3279
T T ci
T T ci
Total
C1
C2
C3
C4
C5
C6
C7
C8
C10
Average
6.3 (65)
4.1 (21)
5.8 (86)
5.2 (57)
2.1 (6)
4.9 (63)
6.1 (21)
0.2 (1)
5.8 (22)
2.1 (6)
2.4 (5)
11.9 (3)
1.8 (1)
13.7 (25)
13.7 (25)
5.0 (11)
5.5 (18)
5.3 (29)
2.1 (6)
2.4 (5)
11.9 (3)
1.8 (1)
5.7 (169)
7.6 (71)
6.2 (240)
i,exp
i,cal
)/H
i,exp
3
i
lni linear
(i )linear
(14)
is iC
T T ci
(15)
ln
pis
1
4.5 1
PCi
Tri
T Tci
(16)
The solid line in Figure 6 stands for Eq. 16. Table 2 includes
the values of (AAD)H for the data covering T Tci. Combining
Eq. 15 and the denition of the acentric factor proposed by
Pitzer,7 the value of the acentric factor for the hypothetical
liquids h, is given as
h 0.16
(17)
Prediction of H
i
3280
December 2005
Using Eq. 18, the H i data for the solute alkane i and the
solvent alkane j are transformed into the H i data for the solute
alkane i and a hexadecane solvent, where Eqs. 4 and 14 are
used for the derivation of the following equation
H i,HD
H iexp
i
(i )linear
0.173
28.3
(Qi,HD Qij )
T
(18)
AIChE Journal
Figure 7. Comparison between experimental and predicted Henrys law constants for a solvent
hexadecane.
(E) 286 points from Henrys law constant data; () 225
points from innite-dilution activity coefcient data; ()
predicted from the model proposed in the present study (Eqs.
4 and 14 16).
p is is i linearexp
H i,HD
i
(i )linear
0.173
28.3
(Qi,HD Qij )
T
(19)
The data from the 225 points for ( i )linear compiled by Kato
H i,HD
predicted from Eqs. 4 and 14 16. The data convergence
December 2005
i iPa
exp
vi0 Pa pis
RT
(20)
E,
i(Pa)
, the partial molar excess enthalpy of solute i, h i(Pa)
was
E,
calculated from ln i(Pa) /(1/T) h i(Pa) /R and the entropy,
E,
E,
E,
s i(Pa)
, from s i(Pa)
/R ln i(Pa)
h i(Pa)
/RT. Figure 8 inE,
E,
cludes the calculated values of s i(Pa) /R and h i(Pa)
/RT. Figure
E,
E,
8 shows that absurd relationships s i(Pa) 0 and h i(Pa)
0 hold
h i(Pa)
and s i(Pa)
were calculated from ln i(Pa)
.
3281
ln iFHMV ln
vi0
vi0
vj0
vj0
(24)
i(Pa)
caused by the decrease in the Poynting factor; therefore,
ln i is convenient for describing the solution structures
around a solute molecule, when the Poynting factor is remote
from unity. Such a discussion is not possible until the accurate
prediction method for H i of simple uids is established.
ln
i predicted from conventional models
In Figure 9, the values of h E,
/RT (F) derived from the
i
inverse temperature derivative of ln i and the corresponding
s E,
/R values (f) are plotted vs. Ni for a hexadecane solvent
i
at 298.15 K; they were calculated as ln i /(1/T) h E,
/R
i
and s E,
/R ln i h E,
/RT. Figure 9 shows a relationi
i
ship, s E,
0, reecting the destruction of the parallel framei
works formed in solvent alkanes by the introduction of a solute
molecule. It also shows that, for a molecule of a simple uid,
the parallel frameworks are retained without destruction because s E,
is almost equal to zero for methane. An entropy
i
enthalpy compensation, 298.15s E,
/h E,
1.8, holds for
i
i
any solutesolvent combination at innite dilution and at
298.15 K.
Innite-dilution activity coefcients dened by the regularsolution theory, ln iReg , are given as follows9
ln iReg vi0 i j 2 /RT
(21)
i H Vi RT/v i0 0.5
(22)
qi
qi
1
qj
qj
(23)
December 2005
ln iFV ln
vi0 v*i
vi0 v*i
1 0
0
vj v*j
vj v*j
(25)
ln i 1 i 2 A 2 B
qi
A i
qj
(26)
AIChE Journal
Nj
Y i, ave *
(AAD)1**
N data
Reference
T ri
3
3
3
2
2
2
1
1
1
10
5
4
10
4
3
6
5
2
0.86
0.81
0.85
0.89
0.82
1.08
0.94
1.05
1.16
14
19
15
15
18
9
17
15
16
8
11
3
8
2
14
16
145
16
1.201.38
0.931.25
0.870.98
1.241.67
0.99
0.640.88
1.632.16
1.461.90
0.901.12
(27)
Table 4. Values of
i Calculated from the Hi Data for Nonpolar Gases and Liquids in Hexadecane at 298.15 K
Solute
Simple uids
Ar
Kr
Xe
CH4
Spherical
CCl4
Nonspherical
Ethane
Pentane
O2
N2
CF4
CO2
a
5
H
i 10 (Pa)*
S i /S methane
i /( i ) alkane
298.15/T ci
354a
121a, 151b
38.9b
166a
0.81
0.93
1.11
1
0.001
0.005
0.008
0.008
0.15
0.67
0.90
1
1.18
0.86, 1.08
0.88
1.09
1.98
1.42
1.03
1.57
0.14b
2.03
0.193
11.91
1.00
0.54
27.6c
1.41b
409a
753a
552a
1.3
2.17
0.81
0.92
1.27
0.099
0.197
0.025
0.039
0.177
9.53
11.33
3.87
5.31
17.36
0.98
0.69
1.93
2.36
1.31
734d, 713e
0.98
0.239
30.41
1.09
0.92
1.48
2.71
19.2
3.62 105,
3.52 105
0.98
*References: Hesse et al. (1996)28; Abraham and Whiting (1990)40; Richon and Renon (1980)20; Hayduk et al. (1972)16; King and Al-Najjar (1977).19
AIChE Journal
December 2005
3283
8.
9.
10.
11.
Conclusions
Universal methods for the prediction of Henrys law constants (H i ) for alkane/alkane binaries have not been established. The present study established a practical prediction
method for the H i values of alkane/alkane binaries representing the H i data in the literature covering alkane solutes from
methane to decane, and alkane solvents from heptane to tetracosane at temperatures from 258 to 475 K. The original ideas
contained in the present study are summarized as follows.
From Eq. 4 combined with Eqs. 14 to 16, Henrys law
constants for alkane/alkane binaries for the temperature ranges
from Tri 3 are accurately predicted, where innite-dilution
activity coefcients for simple uids in alkane solvents are
modied as Eq. 14 and hypothetical liquids are dened in terms
of the vapor pressures given by Eq. 16 and the fugacity coefcients given by Eq. 15 at temperatures above solute critical. If
a pure liquid at the system temperature and at the saturated
vapor pressure is chosen as a standard state of activities, the
innite-dilution activity coefcients based on this standard
state are convenient for representing solution structures and
molecular interactions in alkane/alkane binaries at innite dilution,
where
the
entropy enthalpy
compensation,
298.15s E,
/h E,
1.8, holds. Conventional models, such
i
i
as the regular solution, the FloryHuggins, and the free volume
models, fail to accurately predict the partial molar quantities,
s E,
and h E,
of simple uids. Of the 42 data sets of the
i
i
high-pressure VLE data for alkane/alkane binaries in the literature, the nine data sets listed in Table 3 are in the best
agreement with those predicted by the present model under
dilute conditions, and they are useful for developing a model
describing high-pressure VLE. The present model can predict
the H i values of not only alkanes but also a substance having
a smaller or the same acentric factor with an alkane that has the
same molecular surface area with the substance, where the
measure of the molecular surface area is given by Eq. 27; the
present model accurately predicts Henrys law constants for
Ar, Kr, Xe, and CCl4 in hexadecane.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
Literature Cited
1. Chappelow CC III, Prausnitz JM. Solubilities of gases in high-boiling
hydrocarbon solvents. AIChE J. 1974;20:1097-1103.
2. Rodriguez AT, Patterson D. Prediction of activity coefcients and
Henrys constants at innite dilution for mixtures of n-alkanes. Fluid
Phase Equilib. 1984;17:265-279.
3. Kato S, Hoshino D, Noritomi H, Nagahama K. Determination of
innite-dilution partial molar excess entropies and enthalpies from the
innite-dilution activity coefcient data of alkane solutes diluted in
longer-chain-alkane solvents. Ind Eng Chem Res. 2003;42:4927-4938.
4. Prausnitz JM, Shair FH. A thermodynamic correlation of gas solubilities. AIChE J. 1961;7:682-687.
5. Yen LC, McKetta JJ Jr. A thermodynamic correlation of nonpolar gas
solubilities in polar nonassociated liquids. AIChE J. 1962;8:501-507.
6. Prausnitz JM, Chueh PL. Computer Calculations for High-Pressure
VaporLiquid Equilibria. New York, NY: Prentice-Hall; 1968.
7. Pitzer KS. The volumetric and thermodynamic properties of uids. I.
3284
12.
December 2005
28.
29.
30.
31.
32.
AIChE Journal
33.
34.
35.
36.
AIChE Journal
December 2005
37.
38.
39.
40.
Liquid Equilibria for Mixtures of Low Boiling Substances (Data series). Vol. VI. Frankfurt am Main, Germany: DECHEMA; 1982.
Lhotak V, Wichterle I. Vaporliquid equilibriums in the ethane-nbutane system at high pressures. Fluid Phase Equilib. 1981;6:229-235.
Beranek P, Wichterle I. Vaporliquid equilibriums in the propane-nbutane system at high pressures. Fluid Phase Equilib. 1981;6:279-282.
Blanc CJ, Setier JCB. Vaporliquid equilibria for ethanepropane
system at low temperature. J Chem Eng Data. 1988;33:111-115.
Abraham MH, Whiting GS. Thermodynamics of solute transfer from
water to hexadecane. J Chem Soc Perkin Trans. 1990;2:291-300.
Manuscript received Oct. 12, 2004, and revision received Apr. 4, 2005.
3285
R&D NOTES
Mathematical Development
Consider a column packed with an adsorbent in which a
solute adsorbs irreversibly; the process is supposed to be isothermal; axial dispersed plug ow through the bed is further
D ax
c
q
2c
c
1
2 u0
z
z
t
t
q
c p
1 2 cp
q
r
p
1 p
D epore 2
t
t
t
r r
r
(1)
(2)
ui
dz
dt 1
3286
December 2005
AIChE Journal
1 q0
c0
1 2 2 F F y
0
Pe t2
t
t
1 2 F
F y 0
Pe t
ri
ri
R0
(7)
(8)
rR 0
(9)
1 t
1
(5)
When Pe 3 , this equation gives the common constantpattern condition F y , valid for sufciently long times with
favorable equilibrium in the absence of axial dispersion. It
shows that when axial dispersion is not negligible, such an
assumption is surely not safe.
Introducing the dimensionless variables y q/q0 and Fp
cp/c0, the particle mass balance becomes
(6)
r 2dr 1
3
c0 Fp
y
D
t
R 0 epore q0 r
R0
c 0Ut V 1 t
1
1 q 0V
c0 1 2 Fp
y
r
D epore
t
q0 r2 r
r
3
R 03
r 2ydr
(4)
1 2 dF
F y 0
2Pe dT
R0
(3)
3
R 03
December 2005
Np
1 15
D
R 02 epore
dT
5 1 y 1/3 1
(10)
(11)
3287
The shape of the curve y vs. F suggests a functional dependency of the type
y
These equations must be integrated so that the concentrations F and y are determined. The initial conditions for the
solution should be
T T Bp
F y 0
(12)
1 FdT
TdF
1 F
1 F
(14)
or
1 x3
x3
(15)
x x3
dx
x2 x 1
(16)
where 1 .
The analytical integration of this ordinary differential equation gives
(13)
which enables the determination of the TBp that, by displacement, correctly centers the breakthrough curve around T 1.
Numerical Solution
Equations 10 and 11 were numerically solved using subroutine DGEAR from the IMSL library; the zero initial conditions
were used. Fluid (F) and average solid (y ) concentrations for
the same value of the throughput parameter are represented in
Figure 1 for different values of the parameter ratio Np/Pe. In
fact, if Eqs. 10 and 11 are divided, it is clearly seen that the
relation between the uid and average solid concentrations is a
function of this parameter ratio only. It is worth noting that, for
high values of the Np/Pe ratio, the particles saturate before the
bed, that is, the average concentration inside particles, y ,
reaches the unit value while the concentration in the bed voids,
F, is still increasing. As a consequence, the computer stops
with an error message from the package DGEAR before F 1,
a little after y 1.
3288
December 2005
AIChE Journal
5
151
2x3 3x2
lnx2 x 1
2
2
5 3 1 tan1
2x 1
3
51
2 3
5
2
(20)
N pT
5x2
151
2x 3
lnx2 x 1
2
2
5 3 1 tan1
2x 1
3
(21)
Similarly, the nal time, that is, the value of the throughput
parameter corresponding to the complete saturation of the
particles is obtained from Eq. 20 for x 0 (y 1, F 1/);
if this time also corresponds to F 1 (complete
saturation of the bed)
TF 1
1 x 13 5 x 12
5 5x 12
2x 1 3
(2x 1 3)
2
2
2
x 13
(1 ) (1 ) 3 ln 3
2N p
3
(17)
T Bp 1
15(1 )
2x1 1
ln(x12 x1 1) 5 3 (1 )tan1
2
3
5
2N p
(22)
From Eq. 20, with F given by Eq. 15, the analytical solution
in the form of an F vs. T curve can be obtained. This analytical
response is compared with the numerical one in Figure 4, for
different values of the parameter Np/Pe; the response is made
independent of Np by the chosen representation of Np(T 1)
instead of T as abscissa. The comparison is very good, even for
relatively high values of Np/Pe, showing that the approximation
used is reasonable.
(18)
When , 1 and then x1 0; in this case the value of
C is given simply by
C Np
5 51
2
2 3
(19)
December 2005
3289
t Bp 1
1
1
Pe Ns
(25)
T Bp 1 1 3 ln 3
2N p
3
(23)
N p De pore
1
1 1.61
Np
Pe
esolid
N s D
N s De solid
(26)
Ns
1
Pe
e
In this equation, D
and D e solid are the equivalent (from a
solid
model that lumps solid diffusion and axial dispersion) and the
effective (from a model that explicitly separates the contributions of solid diffusion and axial dispersion) solid diffusivities,
e /R 20 are the equivalent and
respectively; N s and Ns 15D
solid
the real numbers of mass-transfer units by intraparticle (homogeneous) solid diffusion and is the adsorption (ion exchange)
column capacity factor.
(24)
December 2005
AIChE Journal
Conclusions
In this article, the analytical solution of Hall et al. for
irreversible equilibrium with pore diffusion is extended for the
case where axial dispersion is present. This is accomplished by
an approximation that relates the concentration in the uid
phase with the average concentration in the solid phase, given
that the constant-pattern assumption establishing the equality
of these concentrations is no longer valid. The analytical solution obtained is then compared with a numerical solution of the
equations using the Gear method,10 showing that the approximation used is reasonable. When the Peclet number tends to
innity (absence of axial dispersion), the solution consistently
reduces to the solution in Hall et al.1
This new solution is used for the development of an equivalent pore diffusivity, valid when a model that lumps pore
diffusion and axial dispersion is considered. Results in terms of
breakthrough curves from both models are similar for Np/Pe
0.2.
When solid diffusion controls, solutions taken from the
literature also establish an apparent solid diffusivity.
1
Notation
c
c0
cp
Dax
D e pore
e
D
pore
D e solid
e
D
solid
F
Fp
L
Np
N p
Ns
N s
q
q0
q
q
qs
r
ri
R0
AIChE Journal
December 2005
t time, s
T dimensionless throughput parameter {[c0(Ut V)]/[(1
)q0V]}
u0 supercial velocity, m/s
ui interstitial velocity, m/s
w concentration wave velocity, m/s
x auxiliary variable, Eq. 15
y dimensionless adsorbed phase concentration (q/q0)
y dimensionless average adsorbed phase concentration (q /q0).
z axial coordinate in the bed, m
Greek letters
Subscripts
Bp breakthrough point
F nal (at saturation time)
Literature Cited
1. Hall KR, Eagleton LC, Acrivos A, Vermeulen T. Pore- and soliddiffusion kinetics in xed-bed adsorption under constant-pattern conditions. Ind Eng Chem Fundam. 1966;5:212-223.
2. Vermeulen T. Theory of irreversible and constant-pattern solid diffusion. Ind Eng Chem. 1953;45:1664-1670.
3. Cooper RS. Slow particle diffusion in ion exchange columns. Ind Eng
Chem Fundam. 1965;4:308-313.
4. Cooper RS, Liberman DA. Fixed-bed adsorption kinetics with pore
diffusion control. Ind Eng Chem Fundam. 1970;9:620-623.
5. Yoshida H, Kataoka T, Ruthven DM. Analytical solution of the
breakthrough curve for rectangular isotherm systems. Chem Eng Sci.
1984;39:1489-1497.
6. Vermeulen T, LeVan MD, Hiester NK, Klein G. Adsorption and ion
exchange. In: Perry RH and Green D, eds. Perrys Chemical Engineers Handbook. 6th Edition. New York, NY: McGraw-Hill; 1984.
7. Convers A, Deves D. Adsorption des hydrocarbures en lit xe sur
tamis moleculaires. Rev Inst Fran Petrole. 1965;20:1720-1754.
8. Ikeda K, Ohya H, Kanemitsu O, Shimomura K. La courbe de percee
de ladsorption en lit xe pour une isotherme favorable. Chem Eng Sci.
1973;28:227-239.
9. Ozil P, Bonnetain L. Dynamical adsorption in xed bed. Chem Eng
Sci. 1977;32:303-309.
10. Gear CW. Numerical Initial Value Problems in Ordinary Differential
Equations. Englewood Cliffs, NJ: Prentice-Hall; 1971.
Manuscript received Jan. 6, 2005, and revision received Apr. 18, 2005.
3291
Introduction
The majority (90%) of the regulatory loops in the process
industries use conventional proportionalintegral derivative
(PID) controllers. Because of the abundance of PID controllers
in practice and the varied nature of processes that the PID
controllers regulate, extensive research studies have been dedicated to the analysis of the closed-loop properties under PID
controllers and to devising new and improved tuning guidelines
for them, focusing on closed-loop stability, performance, and
robustness1-7 (also see the survey papers8,9). Most of the tuning
rules are based on obtaining linear models of the system, either
through running step tests or by linearizing a nonlinear model
around the operating steady state, and then computing values of
the controller parameters that incorporate stability, performance, and robustness objectives in the closed-loop system.
Although the use of linear models for the PID controller
tuning makes the tuning process easy, the underlying dynamics
of many processes are often highly complex because of such
phenomena as the inherent nonlinearity of the underlying
chemical reaction or operating issues such as actuator constraints, time delays, and disturbances. Ignoring the inherent
nonlinearity of the process when setting the values of the
controller parameters may result in the controllers inability to
stabilize the closed-loop system and may call for extensive
retuning of the controller parameters.
The shortcomings of classical controllers in dealing with
complex process dynamics, together with the abundance of
such complexities in modern-day processes, have been an
important driving force behind the signicant and growing
body of research work within the area of nonlinear process
control over the past two decades, leading to the development
of several practically implementable nonlinear control strateCorrespondence concerning this article should be addressed to P. D. Christodes at
pdc@seas.ucla.edu.
3292
December 2005
AIChE Journal
process behavior explicitly into account. The central idea behind the proposed method is the selection of the tuning parameters in a way that has the PID controller emulate, as closely as
possible, the control action and closed-loop response obtained
under a given nonlinear controller, for a broad set of initial
conditions and set-point changes. To this end, classical tuning
guidelines (typically derived on the basis of linear approximations, running open or closed-loop tests) are initially used in the
rst level to obtain reasonable bounds on the range of stabilizing tuning parameters over which the search for the parameters best matching the PID and nonlinear controllers is to be
conducted. In addition to stability, performance and robustness
considerations for the linearized closed-loop system can be
introduced in the rst level to further narrow down the parameter search range. The bounds obtained from the rst level are
then incorporated as constraints on the optimization problem
solved at the second level to yield a set of tuning parameter
values that enforce closed-loop behavior under the PID controller that closely matches the closed-loop behavior under the
nonlinear controller. Implications of the proposed method, as a
transparent and meaningful link between the classical and
nonlinear control domains, as well as possible extensions of the
tuning guidelines and other implementation issues, are discussed. Finally, the proposed tuning method is demonstrated
through a chemical reactor example.
K c, I, D
uPID t Kc e
0t etdt
de
D
I
dt
et y sp y PID
(1)
December 2005
t final
s.t.
x t f xt g xtu PIDt
y PID h x
1K cc K c 4K cc
2 Ic I 5 Ic
3 Dc D 6 Dc
K c, I, D argmin J
(3)
where ysp is the desired set point; ynl and unl are the closed-loop
process response and control action, respectively, under the
c
nonlinear controller; K cc , Ic , and D
are parameter values obtained using (rst-level) tuning rules based on linear models;
and 0 i 1, i 1, 2, 3 and 1 i , i 4, 5, 6 are
design parameters.
Remark 1. The optimization problem of Eqs. 2 and 3
computes values for Kc, I, and D such that the closed-loop
control action and process response under the PID controller
are similar to those under the nonlinear controller, while being
Vol. 51, No. 12
3293
within acceptable ranges of the values obtained from the classical tuning methods. The method thus allows ne-tuning the
closed-loop performance under the PID controller to mimic
that of the nonlinear controller. Note that, in principle, the PID
controller parameters could have been tuned to mimic any
desired arbitrarily chosen closed-loop behavior. Mimicking the
behavior of the nonlinear controller, however, is a meaningful
objective because the nonlinear controller uses the nonlinear
system dynamics in generating the prescribed input and output
response, and therefore provides a target closed-loop system
behavior that is realizable. Note also that the performance
index of Eq. 2 can be generalized to include a weight factor (in
front of the input penalty term) to quantify the relative importance of the two main terms: the rst one that aims at minimizing the closed-loop output mismatch associated with the
nonlinear vs. the PID controller, and the second input penalty
term. Furthermore, the inclusion of a weight factor would
ensure consistency of units for the two terms in the performance functional.
Remark 2. It should be noted that the controller actually
implemented in the closed-loop system is the PID controller
with xed parameter values and that it may not always be
possible for the PID controller to exactly match the closed-loop
behavior under the nonlinear controller. The purpose behind
introducing the rst-level tuning is twofold: (1) to ensure that
important objectives (such as closed-loop stability) are not
sacriced in the (possibly unsuccessful) quest for a nonlinear
controllerlike behavior (which is accomplished through the
optimization), and (2) to provide a rational way of constructing
a range of the tuning parameter values over which the optimization is performed. With respect to the rst objective, we note
that the essence of the second-level optimization is to try to nd
the best tuning parameter values that make the PID controller
emulate the behavior of the nonlinear controller. However, this
objective should not come at the expense of more overriding
objectives such as closed-loop stability. In particular, if the
optimization problem were to be carried out without imposing
any constraints on the parameter values, the solution may
indeed lead to a closer match, but no longer guarantee that the
PID controller enforces closed-loop stability when implemented. This is one reason why the proposed method includes
the rst level whose purpose, in part, is to make use of existing
methods for PID controller tuning to rst determine the range
of tuning parameters for which closed-loop stability of the
linearized process model under the PID controller is guaranteed. For such a range of values, the PID controller will enforce
local closed-loop stability when implemented on the nonlinear
process.
Remark 3. Having obtained the stabilizing parameter range
and incorporated it as a constraint on the optimization, the
search for the optimal gain then can take place only over this
range. However, the stabilizing range may be too large to
search over and the designer may wish to limit this range
further by incorporating additional performance and robustness
considerations. The use of rst-level methods (which are based
on the use of linear models) provides a rationalalthough not
necessarily uniqueway of constructing an appropriate subrange to work with. For example, if certain robustness margins
can be obtained (and quantied explicitly as ranges on the
tuning parameters) through the use of existing methods based
on the linearized model, these margins can be incorporated as
3294
December 2005
AIChE Journal
Value
Unit
V
E/R
C A0
T A0
H
k0
cp
UA
F
C As
T Rs
T nom
j
0.1
8000
1.0
400.0
2.0 105
7.85 106
1.0
1000.0
1.667 103
0.001667
0.52
398.97
493.87
m3
K
kmol/m3
K
kJ/kmol
s1
kJ kg1 K1
kg/m3
kJ s1 K1
m3/s
kmol/m3
K
K
December 2005
method involves a form of nonlinear optimization, it is expected that, in general, multiple optimal solutions may exist.
Remark 7. The basic idea behind the proposed PID controller tuning methodology, that is, that of tuning the PID
controller to emulate some other well-designed controller that
effectively handles complex dynamics, can be used to develop
conceptually similar tuning methods for PID control of processes with other sources of complexities (besides nonlinearity)
such as uncertainty, time delays, and manipulated input constraints. The logic behind such extensions is based on the
following intuitive parallel: just as a nonlinear controller is a
meaningful guide to be emulated by a PID controller being
implemented on a nonlinear process, a controller that effectively handles constraints, uncertainty, and/or time delays can
also be a meaningful guide to be emulated by a PID controller
that is being implemented on a process with these characteristics. In principle, the extensions can be realized by adequately
accounting for the complex characteristics of these processes
within both levels of the tuning method. For example, for
systems with uncertainty, classical tuning methods that provide
sufcient robustness margins can be used to come up with the
rst-level parameter values. Then a robust nonlinear controller
(see, for example, El-Farra and Christodes10) can be designed
and the closed-loop proles, obtained under the robust nonlinear controller for a sufcient number of realizations of the
uncertainty (which may be simulated, for instance, using random number generators), may be computed. Finally, the parameter values obtained from the rst-level tuning method may
be improved upon by solving an optimization problem that
minimizes the error over the proles in the representative set.
In a conceptually similar fashion, for systems with constraints,
an anti-windup scheme could be used initially to obtain the
rst-level parameter values. A nonlinear controller design that
handles input constraints can then be chosen for the secondlevel optimization. The PID controller tuning guidelines can
then serve to carry over the constraint handling properties of
this nonlinear controller and improve upon the rst-level tuning
methods in two ways: (1) through the objective function, by
requiring the control action and closed-loop process response
under PID control to mimic that under the constrained nonlinear controller; and (2) through the incorporation of the input
constraints directly into the optimization problem. However, it
should be noted that, although such extensions are intuitively
appealing, a detailed assessment and characterization of their
potential requires further investigation.
Remark 8. Note that the derivative part of the PID controller is often implemented using a lter. This feature can be
easily incorporated in the optimization problem by explicitly
accounting for the lter dynamics. Constraints on the lter time
constant f, obtained empirically through knowledge of the
nature of noise in the process, can be imposed to ensure that the
ltering action restricts the process noise from being transmitted to the control action.
Remark 9. To allow for simple computations, approximaTable 2. PI Tuning Parameters
Tuning Method
Kc
IMC
ZieglerNichols
Proposed method
1.81
2.86
5.56
0.403
4.16
0.297
3295
Figure 2. Closed-loop output (a) and manipulated input (b) prole under a linearizing controller (solid line) and a PI
controller tuned using the proposed method (dashed line).
December 2005
T R
F
C C A k 0e E/RTRC A
V A0
F
H E/RT R
UA
T A0 T R
T TR
k0 e
CA
V
cp
cp V j
(4)
where CA denotes the concentration of the species A; TR denotes the temperature of the reactor; Tj is the temperature of the
uid in the surrounding jacket; U is the heat-transfer coefcient; A is the jacket area; V is the volume of the reactor; k0, E,
and H are the preexponential constant, the activation energy,
and the enthalpy of the reaction, respectively; and cp and are
the heat capacity and uid density in the reactor, respectively.
The values of all process parameters are given in Table 1. At
the nominal operating condition of T nom
493.87 K, the
j
reactor is operating at the unique, stable steady state (C As ,
T Rs ) (0.52 kmol/m3, 398.97 K). The control objective is to
implement set-point changes in the reactor temperature using
the jacket uid temperature Tj as the manipulated input, using
a P, PI, or PID controller.
To proceed with our controller tuning method, we initially
design an input/output linearizing nonlinear controller. Note
that the linearizing controller design is used in the simulation
example only for the purpose of illustration, and any other
nonlinear controller design deemed t for the problem at hand
can be used as part of the proposed controller tuning method.
Dening x [CA C As , TR T Rs ] and u Tj T nom
, the
j
Vol. 51, No. 12
AIChE Journal
Figure 3. Closed-loop output (a) and manipulated input (b) prole using IMC tuning rules for PI controller (solid line),
using ZieglerNichols tuning rules (dotted line), and the proposed method (dashed line).
yt L fh x
L gh x
(5)
where Lfh(x) and Lgh(x) are the Lie derivatives of the function
h(x) with respect to the vector functions f (x) and g(x), respectively; , a positive real number, is a design parameter; and
is the set point. Taking the derivative of the output in Eq. 1 with
respect to time, we get y Lfh(x) Lgh(x)u. Substituting the
linearizing control law of Eq. 5, we get y ( y)/.
Under the control law of Eq. 5, the controlled output y
evolves linearly, to achieve the prescribed value of , and the
design parameter is the time constant of the closed-loop
response.
It is well known that when a rst-order closed-loop response,
with a given time constant, is requested for a linear rst-order
process, the method of direct synthesis yields a PI controller.
Note that the relative order of the controller output TR, with
respect to the manipulated input Tj, in the example of Eq. 4 is
also one. Even though the nonlinear controller is a static
controller, and the PI controller is dynamic, both controllers are
capable of generating closed-loop behaviors that are of the
same kind (a linear rst-order response with a prescribed
closed-loop time constant). This encourages the use of a PI
controller and tuning its parameters to achieve the prescribed
rst-order response.
For the purpose of tuning the PI controller, the nonlinear
process response generated under the nonlinear controller, using a value of 0.25, was used in the optimization problem.
An appropriate range for the tuning parameters was derived
from the Kc and I suggested by the IMC-based and Ziegler
Nichols tuning rules (where the parameters Kcu 6.36 and Pu
5.0 are obtained using the method of relay auto tuning).21 In
particular, the constraints on the values of the parameters were
chosen as follows: for a given parameter, the largest and the
smallest values prescribed by the available tuning methods (in
this case the IMC-based and ZieglerNichols) were chosen and
the upper bound on the parameters was chosen as twice the
AIChE Journal
December 2005
maximum value, and the lower bound was chosen as half the
minimum value: 0.9 Kc 5.7 and 0.2 I 8.2. The values
of the parameters, computed using the IMC method, Ziegler
Nichols, and the two-level PI tuning method are reported in
Table 2.
The solid lines in Figures 2a and 2b show the closed-loop
response of the output and the manipulated input under the
nonlinear control of Eq. 5. Note that the value of was chosen
as 0.25 to yield a smooth, fast transition to the desired set point.
The optimization problem was solved approximately, using the
closed-loop process response under the nonlinear controller to
compute e(t), and the objective function included only penalties on the difference between the control actions under the PI
controller and the nonlinear controller (see Remark 9). The
dashed line shows the response of the PI controller tuned using
the proposed optimization-based method. The result shows that
the response under the PI controller is close to that under the
nonlinear controller and demonstrates the feasibility of using a
PI controller to generate a closed-loop response that mimics the
response of the nonlinear controller.
In Figure 3a, we present the closed-loop responses when the
controller parameters computed using the IMC-based tuning
rules and ZieglerNichols are implemented. As can be seen, the
transition to the new set point under the PID controller tuned
using the proposed method (dashed lines) is fastest when
compared to a classical PI controller tuned using IMC tuning
rules (solid line) and ZieglerNichols tuning rules (dotted line).
The corresponding manipulated input proles are shown in
Figure 3b.
We now demonstrate the application of the proposed method
to the same system, but with CA as the controlled variable and
Tj as the manipulated variable. As in the previous case, we
initially design an input/output linearizing nonlinear controller
Kc
IMC-I
IMC-II
ZieglerNichols
Proposed method
678.8
1208.9
2072.0
951.21
0.95
1.00
0.149
0.978
0.0524
0.114
0.028
0.114
3297
Figure 4. Closed-loop output (a) and manipulated input (b) prole under a linearizing controller (solid line) and a PID
controller tuned using the proposed method (dashed line).
Figure 5. Closed-loop output (a) and manipulated input (b) prole using IMC tuning rules I (dotted line), IMC tuning
rules II (solid line), ZieglerNichols tuning rules (dash dotted line), and the proposed method (dashed line).
3298
December 2005
AIChE Journal
Acknowledgments
Financial support from the National Science Foundation (Grant CTS0129571) is gratefully acknowledged.
11.
12.
Literature Cited
13.
AIChE Journal
December 2005
14.
15.
16.
17.
18.
19.
20.
21.
Manuscript received Apr. 26, 2004, and revision received Mar. 30, 2005.
3299
Fuli Wang
Faculty of Information Science and Engineering, Northeastern University, Shenyang, P. R. China
DOI 10.1002/aic.10568
Published online September 23, 2005 in Wiley InterScience (www.interscience.wiley.com).
Introduction
Batch processes play an important role in many industries
for exible manufacturing of high value-added products. Online monitoring of batch processes is of critical importance in
ensuring operation safety and quality consistency. Multivariate
statistical methods for continuous processes, such as principal
component analysis (PCA) and partial least square (PLS), have
been extended to batch process monitoring with some successes, for example, multiway PCA/PLS,1,2 hierarchical
PCA,3,4 and stage-based sub PCA modeling method.5
The above PCA- or PLS-based monitoring methods have
been originally developed for concerning with static rather than
dynamic relationships among process variables. These methods
require implicitly a statistical assumption of batch independence, that is, the variation in the trajectories among the historical reference batches are purely common-caused variation.1
Batch processes, however, are inherently dynamic processes.
Dynamic behaviors may exist not only within a batch run, but
also from batch to batch. For example, in injection molding,
variations in materials and slowly-changing conditions such as
mold temperature can change process characteristics, that is,
process dynamics, in both time and batch-to-batch sense. Two
batch dynamics, namely within-batch dynamics and batch-tobatch dynamics, are needed to describe such processes. In
general, a batch process involves both within-batch and batchto-batch dynamics at the same time. It is desirable to develop
a batch process monitoring scheme that can capture both within-batch and batch-to-batch dynamics simultaneously.
Correspondence concerning this article should be addressed to F. Gao at
kefgao@ust.hk.
3300
December 2005
AIChE Journal
k n 1)]xi, k x1 i, k, . . . , xj i, k, . . . , xJ i, k
The dimension of the 2-D augmented matrix X is (I m
1)(K n 1) mnJ. Its columns are composed of the
current measurements x(i, k), and the lagged process measurements in its ROS. PCA on X can be formulated by
ti, k Xi,k P
(2)
Xm,n
X Xi,k
XI,K
(1)
where
X i,k xi, k, xi, k 1, . . . , xi, k n 1, xi 1, k, xi
1, k 1), . . . , xi 1, k n 1, . . . , xi m
1, k), xi m 1, k 1, . . . , xi m 1,
AIChE Journal
December 2005
x i, k ti, k Pik
(3)
(4)
where P ik is the rst J rows of P, that is, the loadings for x(i,
k). The above modeling procedure, Eq. 2 4, is named 2-DDPCA method, which can be directly used for the following
online batch monitoring.
Two important parameters need to be determined in 2-DDPCA: 2-D autoregressive orders (m, n), and the number of
retained principal components A. Implicitly, the 2-D-DPCA
may be considered as an equivalent to a multivariate 2-D-AR
model; therefore, order estimation methods for 2-D-AR models
can be borrowed for the determination of the 2-D autoregressive orders (m, n). Most well-known 1-D model order determination methods, such as Akaike information criterion (AIC),
minimum descriptive length (MDL) and minimum eigenvalue
(MEV), have been extended for the order estimation for 2-D
AR models.12 In this article, 2-D-AIC is selected, as it is simple
and effective even for processes with signicant cross-correlations. The order pair consisting of the maximum in the two
directions is selected as the 2-D autoregressive orders of 2-DDPCA. As for the number of principal components to be
retained in 2-D-DPCA, cross-validation method13 can be directly used.
i 1, . . . , I; k 1, . . . , K
(5)
3301
(6)
Illustration
2-D-DPCA modeling
The proposed 2-D-DPCA based process monitoring is tested
with a batch process having the following 2-D dynamics
x 1i, k 0.8*x 1i 1, k 0.5*x 1i, k 1
0.33*x 1i 1, k 1 w 1
x 2i, k 0.44*x 2i 1, k 0.67*x 2i, k 1
0.11*x 2i 1, k 1 w 2
x 3i, k 0.65*x 1i, k 0.35*x 2i, k w 3
x 4i, k 1.26*x1 i, k 0.33*x2 i, k w4
(7)
December 2005
AIChE Journal
Figure 4. Ofine batch-to-batch monitoring of the process for a changed correlation structure from
batch 61 by MPCA.
Figure 6. Ofine batch-to-batch monitoring of the process for a small process drift from batch 61 by
MPCA.
Despite the signicant change in the 2-D autoregressive parameters for x 2 , the numerical differences between the new and
normal x 2 , however, only become signicant after 510 batch
runs later, as shown in Figure 3. Prompt detection of this fault
is a challenge for MPCA-based batch monitoring methods.
Figure 4 shows the ofine batch-to-batch monitoring result
using the MPCA model incorporated with batch-to-batch information.7 The fault can be detected only after batch 82 by
SPE monitoring chart, 22 batches after the occurrence of the
fault. The proposed 2-D-DPCA based method, however, can
detect the fault without any delay, as shown in Figure 5. This
prompt fault detection is attributed to the inherent advantages
of the proposed 2-D-DPCA model, as it can model autocorrelation in time and batch at the same time, and the crosscorrelation among process variables can be also effectively
modeled. Any change of correlation (auto- or cross-) in process
data will make the 2-D-DPCA model fail to explain the
AIChE Journal
December 2005
3303
Conclusions
Motivated by developing models to capture batch dynamics
both in time and batch sense, a 2-D-DPCA modeling has been
proposed for online batch process monitoring. Firstly, process
data is augmented to include lagged measurements in both time
and batch directions. PCA is then applied on the 2-D augmented data matrix to capture not only cross-correlations of
process variables, but also autocorrelations in both time and
batch directions. It is shown that the proposed 2-D-DPCA
model can extract all correlated relationships in process data;
while the residuals can provide an effective statistical basis for
online monitoring of 2-D dynamic batch processes.
Acknowledgment
This work is supported in part by a competitive Earmarked Research
Grant of Hong Kong Research Grant Council, under project number
601104.
Literature Cited
1. Nomikos P, MacGregor JF. Monitoring of batch processes using
multiway
principal
component
analysis.
AIChE
J.
1994;40:13611375.
2. Nomikos P, MacGregor JF. Multiway partial least squares in monitoring batch processes. Chemometrics and Intelligent Laboratory Systems. 1995;30:97108.
3. Wold S, Kettaneh N, Tjessem K. Hierarchical multiblock PLS and PC
models for easier model interpretation and as an alternative to variable
selection. J of Chemometrics. 1996;10:463 482.
3304
December 2005
AIChE Journal
Michael R. Ladisch
Weldon School of Biomedical Engineering and Laboratory of Renewable Resources Engineering (LORRE), and Dept. of
Agricultural and Biological Engineering, Purdue University, West Lafayette, IN 47907
Keywords: pathogen capture, bioselective, membrane, Listeria monocytogenes, escherichia coli, antibody immobilization, pathogen detection
Introduction
Membrane ltration is widely used for sterile ltration, owthrough immuno-ltration, virus removal, and microbiological
analysis.1,2,3,4,5,6 Prior work has described membranes that retain one protein over others.7,8 The purpose of this work is to
capture a target species of living cells from a mixed population
of microorganisms being concentrated in the retentate of solution being pushed through a membrane. We report selective
capture of one living microorganism from another by an antibody immobilized on a lysine-modied polycarbonate membrane. Capture of a relatively large microbial entity (L. monocytogenes at 1 m) by a 75 to 150 fold smaller ligand (the
antibody) occurs with high efciency, when a spacer (poly-Llysine) reacted to the membranes surface is derivatized with an
antibody (P66) specic to Listeria. The results show that isolation of the food pathogen L. monocytogenes from E. coli in
less than 2 h is possible.
AIChE Journal
December 2005
3305
Scheme 1
December 2005
AIChE Journal
Results
Fluorescent microscopy from batch experiments show that L.
monocytogenes attaches nonspecically to poly-L-lysine/glutaraldehyde membranes and is not removed by washing three
times with PBS buffer (Figure 2a). Each dot represents a
labeled bacterial cell. BSA covalently bound to the membrane
using poly-L-lysine/glutaraldehyde blocks L. monocytogenes
adsorption13 (absence of dots in Figure 2b). When L. monocytogenes is contacted with P66 membrane, it binds (Figure 2c)
while E. coli does not (Figure 2d). When a mixture of L.
monocytogenes and E. coli was incubated for 40 min with P66
membrane, at 25C, and the membrane was then washed three
times with 1 mL of PBS buffer at 25C to remove nonspeciAIChE Journal
December 2005
3307
Acknowledgments
This research was supported through a cooperative agreement with the
Agricultural Research Service of USDA (Project No. 1935-42000-035),
and National Defense University (Contract No. DABJ29-03-P-0022). The
3308
December 2005
authors would like to thank Dr. J. Paul Robinson and Jennifer Sturgis for
their assistance on the uorescent microscope, and Debra Sherman and
Chia-Ping Huang of the SEM facility at Purdue University for obtaining
the scanning electron micrographs. We thank Dr. Miroslav Sedlak, Dr.
Nathan Mosier, and Dr. Paul Robinson for review of this article; and Dr.
Richard Linton, Director of the Food Safety Engineering Center at Purdue
University, and Randy Woodson, Dean of Agriculture for support of this
work, leadership, and the multidisciplinary environment which makes it
possible.
Literature Cited
1. Peterkin PI, Sharpe AN. Membrane ltration of dairy products for
microbiological analysis. Appl Environ Microbiol. 1980; 39:11381143.
2. Sharpe AN, Peterkin PI, Dudas I. Membrane ltration of food suspensions. Appl Environ Microbiol.1979;37:21-35.
3. Abdel-Hamid I., Ivnitski D, Atanasov P, Wilkins E. Highly sensitive
ow-injection immunoassay system for rapid detection of bacteria.
Anal Chim Acta. 1999;399:99-108.
4. Brewster JD, Mazenko RS. Filtration capture and immunoelectronchemical detection for rapid assay of Escherichia coli O157:H7.
J Immunol Methods. 1998;211:1-8.
5. Mazenko RS, Rieders F, Brewster JD. Filtration capture immunoassay
for bacteria: optimization and potential for urinalysis. J Microbiol
Methods. 1999;36:157-165.
6. Ladisch MR. Bioseparations Engineering: Principles, Practices, and
Economics. New York: Wiley; 2001:220-230.
7. Knudsen HL, Fahrner RL, Xu Y, Norling LA, Blank GS. Membrane
ion-exchange chromatography for process-scale antibody purication.
J Chromatogr A. 2001;907:145-154.
8. Yang H, Viera C, Fischer J, Etzel MR. Purication of a large protein
using ion-exchange membranes. Ind Eng Chem Res. 2002;41:15971602.
9. Chen WT, Hendrickson RL, Huang CP, Sherman D, Geng T, Bhunia
AK, Ladisch MR. Mechanistic Study of Membrane Concentration and
Recovery of Listeria monocytogenes, Biotechnol. Bioeng. 2005;89:
263-273
10. Suye S, Kumon Y, Ishigaki A. Immobilization of glucose oxidase on
poly-(L-lysine)-modied polycarbonate membrane. Biotechnol Appl
Biochem. 1998;27:245-248.
11. Soltys PJ, Etzel MR. Equilibrium adsorption of LDL and gold immunoconjugates to afnity membranes containing PEG spacers. Biomaterials. 2000;21:37-48.
12. Bhunia AK. Antibodies to Listeria monocytogenes. Crit. Rev. Microbiol. 1997;23:77-107.
13. Huang TT, Sturgis J, Gomez R, Geng T, Bashir R, Bhunia AK,
Robinson JP, Ladisch MR. Composite surface for blocking bacterial
adsorption on protein biochips. Biotechno Bioeng. 2003b;81: 618-624.
14. Huang TT, Chang WJ, Atkin D, Gomez R, Bashir R, Mosier N,
Ladisch MR. Microber assisted fabrication of microuidic channels
using polydimethylsilicoxane. AIChE J. 2003a;49:2984-2987.
15. Bhunia AK, Steele PJ, Westbrook DG, Bly LA, Maloney TP, Johnson
MG. A six-hour in vitro virulence assay for Listeria monocytogenes
using myeloma and hybridoma cells from murine and human sources.
Microb. Pathog. 1994;16:99-110.
Manuscript received Mar 12, 2004, and revision received Apr 13, 2005.
AIChE Journal
To the Editor:
In Calculation of Densities from Cubic Equations of State, (April 2002),
Deiters1 reported that Cardanos method
was not the fastest algorithm for nding
the roots of a cubic polynomial, and he
proposed an algorithm based on iteration
that was found to be faster than Cardanos method. My nding was the contrary, as shown later in this letter.
A cubic polynomial to be solved is Eq.
5, the normalized form, in Deiters,1 and
the computer program for Cardanos
method in C/C language is partially
shown below:
p b1 b2*b2/3.;
(1)
Time (s)
Temperature
(K)
Pressure
(atm)
Cardanos Method
Deiters Iterative
Method
300*
320*
340*
350*
360**
370**
11.81
18.51
27.39
32.72
38.68
45.27
0.821
0.821
0.821
0.821
0.751
0.751
1.262
1.261
1.422
1.422
1.302
1.302
rootq 2.*sqrtp/3.;
(8)
x0 rootq*costheta/3. a2;
(9)
x1 rootq*costheta 2.*M_PI/3.
q b0 b1*b2/3. 2.*b2*b2*b2/27.;
(2)
a2 b2/3.;
(3)
d p*p*p/27. q*q/4.;
(4)
(6)
a2;
(10)
x2 rootq*costheta 4.*M_PI/3.
a2;
(11)
Literature Cited
1. Deiters UK. Calculation of densities from
cubic equations of state, AIChE J. 2002;48:
882-886.
2. Walas SM. Phase Equilibria in Chemical
Engineering. Stoneham,
MA: Butterworth Publishers; 1985 (see example 1.13, p. 49).
Paul H. Salim
Dept. of Coiled Tubing Research
& Engineering
BJ Services Company
6620 36th Street S.E.
Calgary, Alberta, Canada T2C 2G4
e-mail: psalim@bjservices.ca
AIChE Journal
December 2005
3309
Reply:
In his letter, Dr. Salim reports that he implemented the new algorithm for solving cubic
equations, but failed to obtain the CPU time
reduction over the Cardano method that had been
mentioned in our paper.
First of all it has to be stated that we fully agree
with the principles of efcient programming laid
down in his letter. His analysis of the Cardano
algorithm is ne, and his lines of C code practically match those in our program. Therefore the
discrepancies of CPU time reductions are probably not the result of different implementations of
the Cardano algorithm.
It may be worthwhile mentioning that using
the C library cbrt(3M) instead of the generic
power function pwr(3M) for computing cube
roots simplies the Cardano program, but
thatat least on the computers available to us
the construct
x 0.0 ?
RT
2RT*
Vm b
Vm2
(3)
b/V m
T T/2T*
p
pb/2RT*
(4)
T
2
1
(5)
x 3 b 2 x 2 b 1x b 0 0
(1)
(2)
3 2 p T p 0
(6)
December 2005
z3
b2 2 b1
b0
z 2 z 30
with
z x
(7)
powx, 1.0/3.0 :
powx, 1.0/3.0
3310
rather then volumes. For example, we implemented the equation of state of van der Waals in
the ThermoC program package1,2 in the following
way1
b1, b2
(8)
but that would require time-consuming evaluations of root functions. For practical purposes,
however, rough estimates for the roots are sufcient; they can be obtained by extracting the
exponents of the oating-point representations of
the bi with the C library function frexp(3M),
dividing them by 2 or 3, and converting the
largest result back to a oating-point representation with ldexp(3M). With the value obtained
this way, Eq. 7 is solved for z, and from this x is
recovered. This algorithm is stable for arbitrary
values of the polynomial coefcients, but it is no
longer faster than Cardanos method.
Finally it should be noted that a further CPU
time reduction can be achieved for the van der
Waals equation by omitting the reduction step
(the division of the polynomial coefcients by a3
to obtain the form of Eq. 1), if reduced properties
are used as in Eq. 6; this saves one oating point
division and three multiplications. The same is
true for the RedlichKwong equation of state and
its variations,3 but not for the PengRobinson
equation.4 This way of saving CPU time applies
to the Cardano method and the iterative algorithm alike.
Literature Cited
1. Deiters UK. A modular program system for
the calculation of thermodynamic properties
of uids. Chem Eng Technol. 2000;23:581584.
2. Deiters UK. ThermoC project homepage.
http://thermoc.uni-koeln.de/index.html.
3. Redlich O, Kwong JNS. On the thermodynamics of solutions. V. An equation of statefugacities of gaseous solutions. Chem Rev.
1949;44:233-244.
4. Peng DY, Robinson DB. A new two-constant equation of state. Ind Eng Chem Fundam. 1976;15:59-64.
Ulrich K. Deiters
Institute of Physical Chemistry
University at Cologne
Luxemburger Str. 116
D-50939 Koln, Germany
e-mail: ulrich.deiters@uni-koeln.de
AIChE Journal