Professional Documents
Culture Documents
Conditional
1987
of Discrete
Fractures
This traceconditioning
is achievedby forcingthe networkgeneratorto alwaysreproducethe observed
traces.Conditioningmightbe a meansof decreasing
the variabilityof the fracturenetworks.A numerical
simulationmodel has been developedwhich is capableof generatinga fracturenetwork of desired
statisticalpropertiesand solvingfor the steadystate flow. On each fracture disc the flow is discretized
statistical
distributions.
Theexamples
leadto thefollowingconclusions.
Largefractures
andhighfracture
densityimpliesgoodconnectivity
in the networks.A highfracturedensityimpliesa smallvariancein the
flow throughthe network.Trace conditioningdecreases
estimationvarianceonly when the fracture
networkconsists
of largefractures.Fracturestatisticscan be estimatedreasonablywell from fracture
traces observed on a wall.
INTRODUCTION
is small for
Field investigationsof flow in fractured rock have clearly the problem analyzed. Interpretations of stochasticmodel redemonstratedthat modelingthe rock as a homogeneouscon- sults are more difficult if the ensemblevariation is large. Antinuum might be a serioussimplification.For example, the derssonet al. [1984] and Anderssonand Thunvik[1986] investhree-dimensionalflow and tracer test by Neretniekset al. tigated the value of conditioning the network with known
[1985] in the Stripa test site in Sweden shows that the flow in fractures,or known parts of fractures.The objectiveof these
the rock can be very unevenlydistributed.In crystallinerock, studieswas to see if the variability of flow and transport
flow is confined to the fractures, whereas the rock matrix is through a fracture network would decreaseif parts of the
almost impervious.Furthermore, there is growing evidence network were given deterministically,based on observations,
that the hydraulicpropertiesin a singlefracturevary strongly and only the unobservedpart was stochastic.They showed
in the fracture plane "channeling"(see,for example,Abelin that the conditioningdoesdecreasethe predictionuncertainty
[1986]). The uneven flow encounteredin crystalline rock is if the fracturesare long and the fracturedensitylow.
Models of flow in fracture networks must be three dimenprobably causedboth by the discretefracture network and the
channeling within a fracture. This paper only considersthe sional in order to be realistic.For example,as is noted by
discretenetwork; the fracturepropertiesare assumedconstant Long et al. [1985], fractureswhich do not connectin a plane
in each fracture. Channeling,which is to be investigatedin cut might connect in space. Consequently,it is not evident
reality,
future studies, may be viewed as a submodel to a discrete how field data, whichare sampledin three-dimensional
shouldbe interpretedin a plane two-dimensionalmodel. Unnetwork model.
Interest in discrete fracture network flow and transport
modelshas grown in the past few years,although until recently thesemodels have only been investigatedin two dimensions
[Long et. al., 1982; Schwartz et al., 1983; Anderssonet al.,
1984; Robinson, 1984; Smith and Schwartz, 1984; Long and
Witherspoon,1985; Anderssonand Thunvik, 1986]. From these
a fractured
rock behaves
As the measurement
and characterization
of all fractures is
rather
flow or transport problem. The fracture network statisticsdeCopyright 1987 by the American GeophysicalUnion.
Paper number 6W4772.
0043-1397/87/006W-4772505.00
to measure
The fracture
and estimate
network
the fracture
network
is viewed as a realization
statistics
of a stochastic
sity. Furthermore,thesedata contain informationon the specific network investigated,which may allow for conditioning
network
realizations
is a measure of the
quality of the field information. In order to evaluatethe fracture network data a computer code has been developedwhich
first generatesfracturesfrom the stochasticmodels and then
solvesfor steady state flow.
FRACTURE
MODEL
to
1877
to the north 0 is
Eachfracturehasa constanttransmissivity
t (m"/s),so that
rg(r)
rg(r)
the flow in the fracture is proportional to the head gradient.
g'(r)=
-(2)
Transmissivitiesare used rather than fracture apertures as no
well-defined relation between these quantities is available for
fracturesin the field. Although experimentsin singlefractures
[e.g., Abelin, 1986] indicate that the fracture transmissivity where r is fracture radius, g'(r) is the radius distribution for
may vary considerablywithin the fracture, this extensionis fractures intersecting a plane, and / is the expected fracture
not included in the presentmodel. Spatially varying fracture radius.
org(r)
dr !a
The
properties
couldbeincluded
in futuremodels.
In the model the fracture transmissivityis a stochasticvariable and is characterizedby a distribution function ti(t) for
each fracture
class. Transmissivities
in different
fractures
cut
which
the fracture
are
makes
with
the
observation
(3)
The mean chord (and trace) length for circleswith radius distribution given by g'(r) is obtained by integrating (3) over all
radii
work parameters.
Fracture centers for each class i follow a Poisson space
E(L) =
E(LIR)a'(R) dR -
(4)
P(N(V), = k) =
e- iqv(iV)k
kl
When
a distribution
(5)
radius is assumed the
1878
ANDERSSON
AND DVERSTORP'THREE-DIMENSIONAL
NETWORKSOF FRACTURES
sampled trace length distribution h(1) is a biased representation off(/) as the full extensionof long fracturescannot
be observedand as long traces have a higher probability of
intersecting the scanline e.g., Baecher and Lanney, 1978;
Priest and Hudson, 1981]. Again, these biases must be dealt
with separatelyfor different distributions.
Estimatesof E(L) by the scanlinemethods require a narrow
classificationwith respectto the angle betweenthe trace and
the scanline;thus only a portion of the sampledpopulation
can be usedin the individual estimates.If the trace intensityis
low and obvioussetswith preferredorientation are absent,the
usefulnessof these methods are severely reduced. Fur-
2Di
'" __
El
-- Nc i=1 Nc
2 ni + mi
i=1
(10)
i=1
Orientation
1982]:
f, -
(6)
2n+m
fracture
orientation.
(7)
for fracture
orientation.
direction
coin-
cides with the pole of the unit sphere, the Fisher distribution
becomes[Mardia, 1972]
tsD
L3--(n+ m+p)s-2D
If the mean
teeKoso sin 0
(8)
f(O, (.p)=
4t sinh c
(11)
where
Thecircular
sampling
region
isnotsuitable
if theheight
h is
small relative to the length of the observationregion as only a
small portion of the trace population contributes to the individual estimates.For this case a better estimate is achieved by
sampling from the maximum number of circles with diameter
h that can be placed in the observation region without one
overlapping another. The sumsfrom each circle of D, n, and m
Fracture Density
ANDERSSON
ANDDVERSTORP.'
THREE-DIMENSIONAL
NETWORKS
OFFRnCTURES
tion areas follows a Poisson distribution with mean
1879
we somewhat unrealistically assumecould be measured exactly), and the position of the trace. Furthermore, the fracture
E(noo
s)= ,gVE
Vr = lim (Po,s(V)
V)
(14)
be expressedas
g(rll)-
g'(r)
1 - G'(I/2)
r > 1/2
g'(r)-
rg(r)
(17a)
g(rll) = 0
f= noos/V
(16)
Once the radius r and the orientation (0, (p) are known, the
remaining step is to determine the position of the fracture
center. If the trace is uncensoredthe fracture could only be
locatedat two positions,insideor outsidethe wall, with equal
probability. If the trace is semicensored
the center is placed
somewhereon an arc in the fracture plane, with radius r and
censoredthe center is placed anywhere in the intersection of
the two circular domains in the fracture plane, with radius r
and origins in the two trace ends.
SOLVER
Generation of Fractures
mating the stochasticnetwork model parameters, as is describedin the previoussection,they also contain information
on specificfractures in the network investigated.A fracture
network model is conditioned by the observedtraces by requiring that all realizationsof this model reproducethe observedtraces and only those. This means of geometrical conditioning used by Anderssonet. al. [1984] and Anderssonand
Thunvik [1986] for networksin two dimensionsis readily exto three-dimensional
(17b)
obtained.
DISCRETE NETWORK
tendable
g'(r) dr
value for M.
CONDITIONING
t'
N
M
G'(r) =
(15) where g(rll) is the radius distribution for the fracture con-
Ve =
r< 0
domains.
whether
it intersects
or is included
in the flow
The generationof the fracturesis split into two parts. First, domain. Only those fracturesthat do so are kept for further
the fractures that produce the observedtraces are generated analysis.
one by one. In this generation the distributions for fracture
Conditioning with observedtraces as describedin the preradii, orientation, and fracture centerposition are conditioned vious section is optional. For the generatednetwork one can
by the information from the observed traces. This is done as also calculate the traces on specifiedobservationareas. Such
follows.In a seconastep the unobservablefracturesare gener- computed trace maps makes it possibleto analyze a synthetated. From the number of observedfracturesnob
s, a fracture ically generated network with the estimation technique decenter density is estimated with (17). From this fracture den- scribed earlier.
sity the total number of fracture centersN in the generation
volume is generated using (1). Of the total of N fractures Forming Fractures Into Flow Network
The lines of intersection with the flow domain boundaries
N-nob
s fracturesare generatedso that they do not intersectthe
observation areas. This generation step is simply based on are calculated for each fracture as well as their intersection
trial and error. First a fracture is generated from the un- lines with the other fractures.Figure 1 is a perspectiveplot of
conditioned
distribution.
this
1880
line.
cmh(xm,
Ym)=
-- m +h n] aS
(20)
makes it possibleto determine whether two fractures are connectedwithout solvingfor the actual flow.
m= In[(x--Xm)
2+ (V--V/)2]
and q is the flux vector out from the boundary, i.e.,
h
q = -t
on an outer boundary
Theouterboundary
of thediscis composed
of circular
arcs,
which define the end of the fracture, and straight lines, which
are the intersection with the box boundaries. The straight lines
inside the disc are the intersections
on an interior line
(22a)
(22b)
V:h =0
(21)
(18)
cmhm
=-j
'=
m
dS
+hj dS (23)
Sj
where h (meters) is the hydraulic head. At the outer boundThe locations of elementson the straight lines arc chosenso
aries of the disc, the boundary conditions are given specified that there arc at least one clement between linc intersections
head (someof the straight lines)and no flow (the arcs and the in the fracture plane. The distance between two linc intersecrest of the straight lines). At the interior lines the hydraulic tions is divided into a large enough number of clementsso
head is the same in both of the intersectingfractures, and the that no element exceedsa specifiedmaximum size. The circular arcs arc approximated with line segmentswhere the segment angledoesnot exceeda presetmaximum angle.
At boundaries with prescribedhead the outward flux is unknown. However, when performing the elimination of nonglobal nodesdescribedin the following sectionit is easierto work
with only heads as unknowns. Voss [1984], and others, have
qB= -- V(hB-- )
(24)
1881
(26)
Class
where{h'} and {q'} are the vectorsof h and q both for the
nodescommonto two fracturesand the nodesat the prescribedheadboundaries.
From (26),{q'} canbe obtainedfrom
n/2
0.0
20.0
n/2
n/2
20.0
0.0
0.0
20.0
(27)
Thesedistributionsare the samefor all setsof input parameters. Fracture radii follow a negativeexponentialdistribution.
Six different combinations(cases)of fracture mean radius t
and fracturedensity; are used(Table 2). The fracturedensity
boundaries.
The resultingsystemof linearequationsis sparse. is chosen(by usingequation (15)) so that the mean number of
No algorithmsfor effectivenumberingis used. Instead the fractures that intersect the face at z = -5 is the same, in this
equationis solvedwith the NAG F01BRF/F04AXF sparse instance 20, for the first three cases and half of that number in
Finally, (27) introducedinto (19) makespossiblethe formation of a global systemof linear equationsin terms of the
hydraulicheadat fractureintersections
and at prescribedhead
mal distribution
elements[Robinson,1986].
SIMULATION EXAMPLES
works in two-dimensionaldomains.However,a true judgement of the practical value of discrete fracture networks
-5.
simulationexamplesand are equalto the data usedto generate the six referencenetworks.The reasonfor prescribingthe
data in this manneris that we want to explorehow well the
fracturestatistics,if they are known, describethe hydraulic
properties of a fracture network. However, in an additional
example the stochastic data are estimated from the observed
physicalprocesses)
in the network.
SpecifiedStatisticalData, No Conditioning
Case It, m
Fig. 3.
1
2
3
4
5
6
4.0
2.0
1.0
4.0
2.0
1.0
Class 2
, m-3
3.0
7.5
1.7
1.5
3.8
8.2
x
x
x
x
x
x
10- 3
10- 3
10- 2
10 -3
10 -3
10 -3
It, m
4.0
2.0
1.0
4.0
2.0
1.0
Class 3
, m-3
3.0
7.5
1.7
1.5
3.8
8.2
x
x
x
x
x
x
10- 3
10- 3
10- 2
10 -3
10 -3
10 -3
It, m
4.0
2.0
1.0
4.0
2.0
1.0
, m-3
1.5 x
3.5 x
8.3 x
7.5 x
1.8 x
4.1 x
10- 3
10- 3
10- 3
10 -'
10 -3
10 -3
1882
-5.00
- 1 0.00
1 0.00
-5.00
- 1 0.00
1 0.00
-10.00
10
-10.00
10,00
O0
-10.00
10.00
-lO.OO
lO.OO
-5.00
-5.00
Fig. 4.
Fracture traces observedon the box face at z = -5 m for (a) case 1, (b) case2, (c) case3, (d) case4, (e) case5, and
(f) case 6.
are Conditioned
Case
1
2
3
4
5
6
#, m
4.0
2.0
1.0
4.0
2.0
1.0
m3/s
1.0
4.2
5.8
5.0
2.0
1.0
x
X
x
x
x
x
10- 2
10 -3
10 -
10 -3
10 -3
10- 3
a/(Q)
0.64
0.60
0.69
0.9
1.1
P.o
0.0
0.0
0.57
0.05
0.14
0.98
With
Face at z =
Case
1
2
2
2
3
4
5
6
m
4.0
4.0
2.0
2.0
1.0
4.0
2.0
1.0
m3/s
2.0
3.6
2.4
4.4
2.8
6.1
5.0
0.0
x
x
x
x
x
x
x
10- 2
10-3
10- 3
10 -3
10 -
10- 3
10 -4
on the
--5 m
m3/s
1.3
1.0
6.6
4.7
1.9
5.5
9.0
8.5
x
x
x
x
x
x
x
x
10- 2
10-2
10- 3
10 -3
10 -3
10- 3
10 -'
10 -4
0.46
0.42
0.93
0.58
0.69
0.78
0.94
0.74
0.0
0.0
0.0
0.0
0.25
0.04
0.40
0.56
ANDERSSON
AND DVERSTORP.'
THREE-DIMENSIONAL
NETWORKSOF FRACTURES
1883
iN
[i!
IE
Fig. 5. Schmidtpolar plotsof the tracesobservedat the box facez = - 5 m for (a) case1, (b) case2, (c) case3, (d) case4,
(e) case5, and (f) case6.
and below
works.
which
almost
no networks
conduct.
One
additional
reference
network
two standard
deviations
each reference
network
50 Monte-Carlo
simulations
used.
variance
in fracture
networks
if the mean
fracture
radius
is
1884
ANDERSSON
AND DVERSTORP:
THREE-DIMENSIONAL
NETWORKSOF FRACTURES
Case
Scanline
#max
#min
#max
#min
(/9
K
2.42
4.12
3.75
4.26
2.16
1.39
5.67
1.61
1.31
2.29
1.27
1.53
5.37
1.85
1.34
1.11
1.53
1.07
1.57
0.94
3.32
4
5
6
7.50
3.50
1.98
5.97
0.82
1.49
4.28
1.31
1.81
1.43
1.08
1.24
1.32
1.49
1.47
5.42
4.41
5.00
1.92
4.95
2.23
TABLE
7.
Mean
Fracture
000
of Observed
#, m
Hob
s
1
2
3
3.93
1.46
1.22
20
17
16
2, m3
6.94 x 10-3
2.26 x 10 -2
2.50 x 10-2
the trace angle to the scanline [see Priest and Hudson, 1981].
A pooled estimate of E(L) was then obtained by weighting
Maximum #maxand minimum #minmean radius were estimated with the number of traces in each interval.
with the circular sampling region and with the scanline method.
If the fracture density is high the estimates of the mean
Mean orientation (0, o)and in Fisher distribution.
intersect
the observation
area
all ensemblevariance remains. In practice, it is clear that conditioning of fracture networks cannot only be based on pure
geometricalinformation. Hydraulic test and tracer test have to
fracture radius lie relatively close to the real values, but when
the fracture density is low the reliability of the estimatescould
be questioned. At any rate, a large difference between the
different estimatorsindicatesif the radius estimate is poor. A
low value is estimatedfor the cparameter in the Fisher distribution. This low value indicates almost random directions,
be utilized.
three different
Networks
In a final set of examplesit is investigated how well distribution parameters for fracture radius, orientation, and density
can be estimated
real values.
c in the Fisher
From the Observed Fracture
Traces
fracture
on real fractures
is avail-
distribution
orientation
of the trace
The fracture density 9was estimated from (14) by introducing the estimated fracture statistics and the number of observed traces. The mean of the three different # estimatesfrom
the circular sampling method was chosento be the estimate of
the radius mean. Table 7 provides the chosen values of # and
the resultingfracture densities.A singleclassof fracture orientations was assumed.Table 8 showsthe resulting fracture den-
able. When making the estimatesit is assumedthat the distributional forms for these properties are known. This is, of
course, still an idealized situation. However, no a priori assumptionson the number of classesare made.
Figures 5a-5f show a Schmidt plot (i.e., the normals of the
sities if three classes of orientation were assumed. The first two
fractures plotted on stereographicprojection) of the observed
of these classesare the ones given in Table 6, whereas the
traces from the six different reference networks.
Orientation
class 1 and class2 but not class 3 (fractures almost parallel to third classis the one specifiedin Table 2. The estimateddenthe observation area) in Table 1 are readily identified from sitiesagreerelatively well with the real valueslisted in Table 2.
theseplots.
In Table 5 appears several estimates of the fracture mean
radius # and estimatesof the mean orientation (0, 6o)and cin
the
Fisher
distribution.
There
it is assumed
that
all
traces
TABLE
6.
Estimated
Fracture
Orientation
Statistics
Traces are Classified Into Three Classes
Class 1
Class 2
When
the
1
2
1.44
1.56
6.21
6.13
28.3
14.0
1.44
1.54
4.76
4.84
24.2
19.0
1.52
0.08
13.7
1.57
1.54
17.9
0.0
0.0
0.0
0.0
were confined
Table 9 and 10 list the estimated mean flow (Q), the ratio
Class 3
Case
fracture
Class 2
Class 3
Case
nob
s
nob
s
nob
s
1
2
3
10
8
8
2.8 x 10 -3
8.8 x 10 -3
1.2 x 10 -2
9
8
8
3.3 x 10 -3
9.8 x 10 -3
1.1 x 10 -2
1
1
0
1.3 x 10 -3
4.0 x 10 -3
0.0
ANDERSSON
AND DVERSTORP:THREE-DIMENSIONAL
NETWORKSOF FRACTURES
1885
NonconductingNetworksPnofor EstimatedFractureStatistics
AssumingThat All FracturesBelongto One Class,No Conditioning
Flow (Q),
Case
m3/s
1
2
3
a/(Q)
8.4 x 10 -3
1.2 x 10- 3
1.0 x 10- 3
Pno
0.66
0.51
0.81
0.0
0.28
0.59
Qreal,
(Q),
Case
m3/s
m3/s
o-/(Q)
1
2
3
2.0 x 10- 2
2.4 X 10 -3
2.8 x 10 -4
1.2 x 10- 2
3.4 X 10 -3
2.1 x 10 -3
0.56
0.63
1.10
Pno
0.0
0.05
0.18
fracture networks.
wereperformed
assuming
a singleclassOffractureorienta- tracer
tions. The results are listed in Table 11. From the table can be
characters
CONCLUSIONS
Qualitatively, there are many similaritiesbetweendiscrete worthwhile to continue research on three-dimensional discrete
network models. However, in a continued investigation it
domains. It is
the combined effect of the fracture radii and fracture density would be necessaryto start to interpret real field data.
One of the moreimportantdirections
of futureresearch
is
that will determine the uncertaintyin flow through a discrete
to
explore
the
validity
of
the
model
assumptions
and
preferfracture network. Large fractures implies good connectivity
(and high conductivity).A high fracturedensityimpliesgood ably subsitutethem with more realisticones.Of the more
it is probablythe onesconcerningspaconnectivity and a small variance in the estimated flow. A important assumptions
fracture networks in two- and three-dimensional
network
with
smaller
fractures
will
more
seldom
conduct
Nonconducting
NetworksPofor EstimatedFractureStatistics
AssumingThat The FracturesBelongto Three Different
Orientation Classes,No Conditioning
(Q),
Case
m3/s
a/(Q)
1
2
3
1.1 x 10 -2
1.5 x 10- 3
9.8 x 10 -4
0.61
0.54
1.01
Pno
0.0
0.23
0.41
1886
REFERENCES
by L. O. Werme,MaterialResearch
Society,Pittsburg,Pa., 1985.
Panek, L. A., Estimating fracture trace length from censoredmeasurementson multiple scanlines,paper presentedat Proceedingsof The
Technol., Stockholm, Sweden, 1986.
International Symposiumon Fundamentals of Rock Joints, SweAndersson,
J., and R. Thunvik,Predictingmasstransportin discrete
dish Natl. Sci. Res. Counc., Bjbrkliden, Sweden,1985.
fracture networks with the aid of geometrical field data, Water
Priest, S. D., and J. A. Hudson, Estimation of discontinuity spacing
Resour.Res.,22(13), 1941-1950, 1986.
and trace length using scanline surveys,lnt. J. Rock Mech. Min.
Andersson,J., A.M. Shapiro, and J. Bear, A stochasticmodel of a
Sci., 18, 183-197, 1981.
fractured rock conditioned by measured information, Water
Robinson,P. C., Connectivity,flow and transportin network models
Resour. Res., 20(1), 79-88, 1984.
of fractured media, AERE Rep. TP1072, Theor. Phys. Div., HarBaecher, G. B., and N. A. Lanney, Trace length biasesin joint surwell, Oxfordshire, United Kingdom, 1984.
veys,in Proceedinosof the 19th U.S. Symposiumon Rock Mechanics,
Robinson, P. C., Flow modelling in three dimensionalfracture netpp. 56-65, American Institute of Mining Engineers, New York,
works, AERE Rep. Rl1965, Theor. Phys. Div., Harwell, Oxford1978.
shire,United Kingdom, 1986.
Brebbia, C., The Boundary Element Method for Engineers, Pentech,
Santalo, L. A., Integral Geometryand GeometricProbability,AddisonLondon, 1978.
Wesley, Reading, Mass., 1976.
Elsworth, D., A hybrid boundary element finite element analysisproSchwartz, F. W., L. Smith, and A. S. Crowe, A stochastic analysis of
cedure for fluid flow simulations in fractured rock masses, lnt. J.
macro dispersion in fractured media, Water Resour. Res., 19(5),
Numer. Anal. Method Geomechan.,10, 569-584, 1986.
Jensen,E. B., Design- and model-basedstereologicalanalysisof arbitrary shapedparticles,Res. Rep. 137, Dep. of Theor. Stat., Univ. of
Aarhus, Denmark, 1985.
645-658, 1983.