Hydraulic Conductivity in Spatially Varying Media

- A Pore-Scale Investigation

J. Jang1, G.A. Narsilio2 and J.C. Santamarina3

Submitted to Geophysical Journal International

November 2009

1
Jaewon Jang
School of Civil and Environmental Engineering,
790 Atlantic Drive
Georgia Institute of Technology
Atlanta, GA 30332, USA
jjang6@gatech.edu; 404-514-2990; Corresponding author
2
Guillermo A. Narsilio
Department of Civil and Environmental Engineering
The University of Melbourne
Parkville, VIC 3010, Australia
3
J. Carlos Santamarina
School of Civil and Environmental Engineering,
790 Atlantic Drive
Georgia Institute of Technology
Atlanta, GA 30332, USA
 1 Hydraulic Conductivity in Spatially Varying Media
2 - A Pore-Scale Investigation
3
4 J. Jang1, G.A. Narsilio2 and J.C. Santamarina3
5
6
7 Abstract
8 The hydraulic conductivity can control geotechnical design, resource recovery, and waste
9 disposal. We investigate the effect of pore-scale spatial variability on flow patterns and hydraulic
10 conductivity using network models realized with various tube size distributions, coordination
11 number, coefficient of variation, correlation and anisotropy. In addition, we analyze flow
12 patterns to understand observed trends in hydraulic conductivity. In most cases, the hydraulic
13 conductivity decreases as the variance in pore size increases because flow becomes gradually
14 localized along fewer flow paths; as few as 10% of pores may be responsible for 50% of the total
15 flow in media with high pore-size variability. Spatial correlation reduces the probability of small
16 tubes being next to large ones, and leads to higher hydraulic conductivity while focused fluid
17 flow takes place along interconnected regions of high conductivity. A pronounced decrease in
18 tortuosity is observed when pore size and spatial correlation in the flow direction are higher than
19 in the transverse direction. These results highlight the relevance of grain size and formation
20 history dependent pore size distribution and spatial variability on hydraulic conductivity, related
21 geo-process, and engineering applications.

1
22 1. Introduction
23 The hydraulic conductivity k depends on the size of pores, their spatial distribution and
24 connectivity. These pore-scale characteristics are defined by grain size distribution and formation
25 history. In turn, hydraulic conductivity controls fluid invasion, flow rate and pore fluid pressure
26 distribution. Consequently, hydraulic conductivity affects storativity, effective stress and
27 mechanical stability, plays a critical role in geotechnical design, determines contaminant
28 migration and the selection of remediation strategies, defines the limits for resource recovery (oil
29 production and residual oil saturation, gas extraction from hydrate bearing sediments, methane
30 recovery from coal bed methane, non-isothermal fluid flow in geothermal applications), and is a
31 central parameter in the design of waste disposal strategies, from nuclear waste to CO2
32 sequestration.
33
34 In this study, we investigate the effect of pore-scale spatial variability on macroscale hydraulic
35 conductivity using network models, following the pioneering work by Fatt (1956a, 1956b,
36 1956c). The main advantage of network models resides in their ability to capture pore-scale
37 characteristics within a physically sound upscaling algorithm to render macroscale properties
38 relevant to the porous medium. Networks can be generated either by assuming an idealized
39 regular geometry, by adopting physically representative networks that capture the porous
40 structure [Bryant et al., 1993] or by mapping the pore structures measured by high resolution
41 tomographic technics onto a network structure [Dong and Blunt, 2009, see also Al-Raoush and
42 Wilson, 2005; Narsilio et al., 2009]. Network model results are consistent with experimentally
43 obtained values of permeability [Al-Kharusi and Blunt, 2007; Al-Kharusi and Blunt, 2008]. The
44 approach has been used to upscale a wide range of pore-scale phenomena such as viscous drag,
45 capillarity, phase change (e.g., ice or hydrate), and mineral dissolution. Consequently, network
46 models have been used to study multiphase flow [Valvatne, 2004; Al-Kharusi and Blunt, 2008],
47 wettability effects in multiphase flow [Suicmez et al., 2008], fine migration and clogging
48 [Kampel et al., 2008], mineral dissolution [Hoefner and Fogler, 1988; Fredd and Fogler, 1998],
49 pressure-induced pore closure [David, 1993], CO2 sequestration [Kang et al., 2005], liquid or gas
50 diffusion through porous media [Laudone et al., 2008; Mu et al., 2008], drying and unsaturation
51 [Prat, 2002; Surasani et al., 2008], the effect of flow localization on diffusion [Bruderer and
52 Bernab, 2001], and resource recovery such as methane production from hydrate bearing

2
53 sediments [Tsimpanogiannis and Lichtner, 2003; Tsimpanogiannis and Lichtner, 2006].
54 Furthermore, pore-scale network models have been coupled to continuum models to conduct
55 field-scale simulations of complex processes such as clogging, reactive flow, and non-Darcian
56 flow near well-bores [Balhoff et al., 2007].
57
58 The first part of the manuscript summarizes previous studies. Then, we provide a detailed
59 description of the numerical model, report statistical results in terms of equivalent hydraulic
60 conductivity, and compare trends against known and analytically derived lower and upper
61 bounds.
62
63 2. Variability in Hydraulic Conductivity Previous Studies
64 Hydraulic conductivity can vary by more than ten orders of magnitude, from very low values in
65 montmorillonitic shale to high values in gravels and boulders. Hydraulic conductivity varies
66 widely even for a given material. The coefficient of variation, defined as the ratio between the
67 standard deviation and the mean, can range from 100% to 800% for both natural sediments
68 [Albrecht et al., 1985; Cassel, 1983; Duffera et al., 2007; Libardi et al., 1980; Warrick and
69 Nielsen, 1980] and remolded sediments [Benson, 1993; Benson and Daniel, 1994]. Data are
70 typically log-normal distributed so that x=log(k/[k]) is Gaussian, where [k] captures the
71 dimensions of k. [Freeze, 1975; Hoeksema and Kitanidis, 1985].
72
73 The correlation length L is the distance where the spatial autocorrelation decays by 1/e0.368.
74 The correlation length for hydraulic conductivity ranges from less than a meter to hundreds of
75 meters. It is typically longer in the horizontal plane than in the vertical direction, in agreement
76 with layering and weathering patterns [Bjerg et al., 1992; Ditmars et al., 1988; DeGroot, 1996;
77 Lacasse and Nadim, 1996].
78
79 The equivalent hydraulic conductivity keq of spatially varying media reflects the distribution of
80 individual values ki, their spatial correlation and flow conditions. Available close-form solutions
81 are summarized in Table 1. In particular, the equivalent hydraulic conductivity keq is (1) the
82 harmonic mean of individual ki values in one-dimensional systems, (2) the geometric mean in
83 two-dimensional media, and (3) higher than the geometric mean when seepage in 3D systems

3
 84 can take place through multiple alternative flow paths. The equivalent hydraulic conductivity in
85 these three cases can be computed in terms of the geometric mean kg and the variance 2 in
86 log(k/[k]), as captured in the following expressions [Gutjahr et al., 1978; Dagan, 1979 - refer to
87 Table 1]:
88
[ (
keq = k h = k g 1 logk
2
2 )] (1D system) (1)

k eq = k g (2D system) (2)

[ (
keq = k g 1 + log
2
k 6 )] (3D system) (3)
89 More complex systems have been studied using equivalent continuum numerical methods. Those
90 results show that (1) flow rate decreases as the coefficient of variation COV(k) increases, and (2)
91 the mean hydraulic conductivity in correlated fields is higher than in uncorrelated fields with the
92 same coefficient of variation COV(k) [Griffiths and Fenton, 1993; Griffiths et al., 1994; Griffiths
93 and Fenton, 1997].
94
95 3. Network models
96 Network models consist of tubes connected at nodes, and can be used to simulate fluid flow
97 through pervious materials. Volume can be added at nodes to reproduce various conditions
98 [Blunt, 2001; Reeves and Celia, 1996; Acharya et al., 2004]. The flow rate through a tube q
99 [m3/s] is a function of fluid viscosity [Ns/m2], tube radius R [m], tube length L [m], and
100 pressure difference between end nodes P [N/m2]:
R 4
q= P = P tube equation Poiseuille (4)
8 L

101 where =R4/8L under isothermal condition, constant viscosity, and constant tube radius.
102 Mass conservation requires that the total flow rate into a node equals the total flow rate out of the
103 node:

q i =0 node equation (5)

104 Equations 4 and 5 can be combined to determine the pressure at a central node Pc as a function of
105 the pressure at neighboring nodes Pi.

4
 Pc =
P
i i
(6)
i

106 If all -values are equal, Equation 6 predicts Pc = (Pa+Pb+Pr+Pl)/4. It is worth noting that this
107 equation is identical to the first order central finite difference formulation of Laplaces field
108 equation.
109
110 Equation 6 is written at all internal nodes to obtain a system of linear equations which can be
111 captured in matrix form:
AP =B (7)

112 where the matrix A is computed with tube conductivities , P is the vector of unknown pressures
113 at internal nodes, and the vector B captures known boundary pressures. The vector P can be
114 recovered as P=A-1B. Once fluid pressures Pi are known at all nodes, the global flow rate Q
115 through the network is obtained by adding the flow rate q (Equation 4) in all tubes that cross a
116 plane normal to the flow direction. The equivalent network hydraulic conductivity in the
117 direction of the prescribed external pressure gradient is calculated from the computed flow rate Q
118 and the imposed pressure gradient between inlet and outlet boundaries. Insightful information is
119 gained by analyzing prevailing flow patterns within the networks as will be shown later in this
120 manuscript.
121
122 Network generation. Networks are realized with pre-specified statistical characteristics. We
123 control the coefficient of variation in tube size, spatial correlation, and isotropy to generate
124 networks with different tube size distribution (mono-sized, bimodal, or log-normal distributed)
125 spatially uncorrelated or correlated and isotropic or anisotropic. Every realization is identified
126 according to these three qualifiers.
127
128 Pore size R is log-normally distributed in sediments. Mercury intrusion porosimetry data for a
129 wide range of soils and effective stress conditions show that the standard deviation in (ln
130 (R/[m])) is about 0.40.2 [Phadnis and Santamarina, 2010]. Examples of statistical distributions
131 used in this study are shown in Figure 1. Throughout the manuscript, the log-normal distribution
132 of pore cross sectional area is used in terms of R2, i.e., log(R2/[R]2) where [R] indicates unit of R.

5
133 Tube R2 values are generated as R2=10a where a is a set of Gaussian distributed random numbers
134 with given standard deviation. Values R2 are scaled to satisfy the selected mean value. While we
135 assume log-normal distribution for network generation, we analyze results and global trends in
136 terms of the mean and standard deviation of R2, that is (R2) and (R2) for each realization.
137 The computer code is written in MATLAB. The run time for each realization in a 2.4GHz
138 processor is ~40 min. The reported study was conducted using a stack of dual-core computers.
139
140 4. Studied Cases - Numerical Results
141 Network models are used herein to extend previous studies on the effect of spatial variability and
142 anisotropy on hydraulic conductivity. Numerical results are presented next. Simulation details
143 are listed in the corresponding figure captions.
144
145 Bimodal distribution - Effect of coordination number and bounds
146 Consider a bimodal distribution made of large and small tubes of relative size RL/RS=5.62 so that
147 their conductivity ratio is kL/kS=103 for constant tube length (refer to Equation 4). Twenty
148 spatially randomly arranged networks are generated for each fixed fraction of small tubes. Two-
149 dimensional networks with coordination number cn=4, 6, and 8 and three-dimensional networks
150 with coordination number cn=6 are used to investigate the effect of coordination number on flow
151 conditions. Computed hydraulic conductivities are averaged for the 20 realizations and plotted in
152 Figure 2 where the mean value is normalized by the hydraulic conductivity of the network model
153 made of large tubes only, kmix/kL.
154
155 Results in Figure 2 show that network conductivity range in three orders of magnitude from
156 kmix/kL=0.001 to 1.0 in agreement with the size ratio (RL/RS)4=103. Hydraulic conductivity values
157 increase as the coordination number increases. There is a pronounced decrease in flow rate
158 where large tubes cease to form a percolating path. Percolation thresholds (readily identified in
159 linear-linear plots - see also Hoshen and Kopelman, 1976) decrease as coordination numbers
160 increase; results are consistent with reported percolation thresholds for various networks: 2D-
161 honeycomb (fraction of small tubes=0.65), 2D-square (0.5), 2D-triangular (0.35), and 3D-simple
162 cubic arrangement (0.25) [Sahimi, 1994; Stauffer and Aharony, 1992].
163

6
164 Analytical solutions for equivalent hydraulic conductivity, and lower and upper bounds
165 summarized in Table 1 are compared to numerical results in Figure 3. The normalized mean
166 hydraulic conductivity for 20 realizations using 2D networks with cn=4 follows the Matherons
167 mixture model. All simulation results are between Wieners and Hashin and Shtrikmans upper
168 and lower bounds (Table 1). Hashin and Shtrikman bounds incorporate the dimensionality of the
169 system resulting in 3D bounds that are shifted toward high keq values compared to 2D bounds.
170 Overall, numerical and analytical results point to higher value of hydraulic conductivity with a
171 larger number of alternative flow paths.
172
173 Coefficient of variation in random networks
174 We explore next the effect of variance in R2 by creating 2D and 3D networks with the same
175 nominal mean (R2). Network statistics, mean (R2), standard deviation (R2), and coefficient of
176 variation COV(R2) are evaluated for each realization. Note that R2-distributions are skewed
177 toward higher values as the coefficient of variation increases (Figure 1-b) even though they all
178 have the same (R2).
179
180 The conductivity of a given realization kdist is normalized by the conductivity kmono of the
181 network made of all equal size tubes, i.e., R2=(R2) and COV(R2)=0. The normalized hydraulic
182 conductivity kdist/kmono decreases as the coefficient of variation of R2 increases (Figure 4) (see
183 similar results in Bernab and Bruderer, 1998). The normalized arithmetic, geometric, and
184 harmonic means computed for each network are shown as shaded areas on Figure 4. The range in
185 normalized hydraulic conductivities for 2D cn=4 networks coincides with the shaded band of
186 geometric means computed for all networks. Computed hydraulic conductivity values for 3D
187 cn=6 and 2D cn=6 networks are the same as the range obtained using Equation 3 and confirm the
188 applicability of the close-form solutions.
189
190 These trends result from the increased probability of large tubes becoming surrounded by smaller
191 tubes, i.e., there is an increased probability of finding a small tube along every potential flow
192 path with increasing coefficient of variation COV(R2). This effect is more pronounced when the
193 coordination number decreases because there are fewer alternative flow paths; in other words,
194 network models with high coordination number are less sensitive to variation in pore size

7
195 COV(R2) because a higher number of alternative flow paths develop in high connectivity
196 condition. Flow patterns are analyzed in detail later in this manuscript.
197
198 Anisotropic, uncorrelated networks
199 When tubes parallel to the predominant fluid flow direction are mono-sized RP (parallel tubes),
200 the flow rate is proportional to RP4, and the distribution of tube size transverse to flow direction
201 RT (transverse tubes) does not affect the global flow rate because there is no local gradient or
202 fluid flow transverse to the main flow direction. This is not the case when tubes parallel to the
203 flow direction are of different size, i.e., not mono-sized.
204
205 Lets consider log-normal distributions for the size RP2 and RT2 of both parallel and transverse
206 tubes. We select different mean values (RP2)(RT2) and adjust standard deviations (RP2) and
207 (RT2) so that both parallel and transverse tubes have the same coefficient of variation COV(R2).
208
209 Results in Figure 5 show that the normalized hydraulic conductivity decreases as the coefficient
210 of variation COV(R2) increases when (RP2)/(RT2)>1. However, the hydraulic conductivity may
211 actually increase when transverse tubes are of high conductivity as shown by the (RP2)/(RT2)=
212 10-2 case: fluid flows along transverse tubes until it finds parallel tubes of high conductivity,
213 mostly with RP2>(RP2).
214
215 Two extreme networks of series-of-parallel and parallel-of-series tubes provide upper and
216 lower bounds to the numerical results (shown as lines in Figure 5). When the ratio of (RP/RT)2 is
217 larger than 10-1, the network responds as a parallel combination of tubes in series. When the ratio
218 of (RP/RT)2 is smaller than 10-1, pressure is homogenized along the relatively large transverse
219 tubes, as captured in the series-of-parallel bound.
220
221 Spatial correlation in pore size Isotropic networks
222 Spatial correlation in pore size upscales to the macroscale hydraulic conductivity in unexpected
223 ways. The methodology followed in this study starts with a set of tubes with fixed (R2) and
224 COV(R2). Then, we use the same set of tubes to generate 100 randomly redistributed spatially
225 uncorrelated networks and other three sets of 100 isotropically correlated networks with

8
226 correlation lengths L/D=5/39, 15/39, and 30/39 (where L is correlation length and D is the
227 network size transverse to the overall flow direction) and for different COV(R2). We use the
228 method by Taskinen et al. (2008) to create correlated fields.
229
230 Hydraulic conductivities are numerically computed for all networks kcor. For comparison, the
231 hydraulic conductivity kmono is evaluated for a network of equal size tubes, i.e., COV(R2)=0. The
232 normalized mean hydraulic conductivity kcor/kmono computed using the 100 realizations for each
233 COV(R2) is plotted versus COV(R2) in Figure 6-a. The normalized mean conductivity decreases
234 with COV(R2) in all cases in agreement with Figure 4, but it is higher in correlated than in
235 uncorrelated networks. Note that the variance from the mean trend also increases with COV(R2)
236 and it is exacerbated by spatial correlation L/D (Figure 6-b).
237
238 Spatial correlation in pore size Anisotropic networks
239 In order to gain further insight into the previous results, we study the effect of anisotropy in
240 correlation length following a similar approach, but in this case we distinguish the correlation
241 length parallel to the overall flow direction LP from the correlation length transverse to the
242 overall flow direction LT. The isotropic case is created with LP/D=LT/D=2/39 so that LP/LT=1.0.
243 High correlation parallel to the flow direction is simulated by increasing LP/D, while high
244 correlation transverse to the flow direction is imposed by increasing LT/D. The study is repeated
245 for three sets of R2 with COV(R2)=0.5, 1.4, and 2.8.
246
247 Average hydraulic conductivity values (based on 20 realizations) are normalized by the hydraulic
248 conductivity kmono of the network made of equal size tubes. Results in Figure 7 show that the
249 normalized hydraulic conductivity kCOV>0/kmono increases as spatial correlation parallel to the
250 flow direction LP/LT increases and it may even exceed the conductivity of the mono-sized tube
251 network in highly anisotropic networks with very high LP/LT values. Otherwise, variation in tube
252 size COV(R2) has a similar effect reported previously: an increase in COV(R2) causes a decrease
253 in hydraulic conductivity (see Figure 4 and 6). Overall, results in Figure 7 point to pore-scale
254 flow conditions similar to those identified in Figure 5.
255

9
256 5. Discussion
257 Numerical results show the evolution of percolation in bimodal system (Figure 2 and 3), and the
258 decrease in hydraulic conductivity with increasing variance in pore size while the mean value of
259 pore size remains constant. This is observed for all types of network topology (Figure 4), and in
260 both spatially correlated and uncorrelated networks (Figure 6). The only exception to this trend is
261 found in highly anisotropic porous media in the direction that favors fluid flow (Figure 5 and 7).
262
263 To facilitate the visualization of flow patterns, we compute tube flow rates (Equation 4), and
264 represent tubes with lines of thickness proportional to flow rate (additional plotting details are
265 noted in figure captions). Figure 8-a shows flow patterns in bimodal distribution networks made
266 of different fractions of small tubes. Flow localizes along dominant flow channels when the
267 fraction of small tubes is 50% which is near the percolation threshold for this network (2D cn=4).
268 Few flow paths are responsible for the global conductivity in networks where the fraction of
269 either small or large tubes is ~50% (Figure 8-a ); conversely, multiple flow paths contribute to
270 the global conductivity in networks made of a majority of either small or large tubes (Figure 8-a
271 and ).
272
273 The fraction of parallel tubes which conducts 50% of the total flow, tubes50%, quantifies this
274 observation (Figure 8-b). The values is tubes50%=50% when all tubes are of the same size, either
275 large or small, which means flow is homogeneous. Fluid preferentially flows along the large
276 tubes so that large tubes are responsible for 50% of the total flow until the fraction of small tubes
277 exceeds ~65%. The participation of small tubes starts to increase above the large-tube
278 percolation threshold (0.5 - Point in Figure 8-b). In general, flow always seeks the larger tubes.
279
280 Distributed tube diameters exhibit a similar response. Most parallel tubes contribute to total flow
281 when the coefficient of variation of R2 is low (Figure 9-a ). Flow becomes gradually localized
282 as COV(R2) increases and fewer channels contribute to global flow (tubes50% in Figure 9-c).
283 Consequently, hydraulic conductivity decreases as shown earlier (Figures 4 and 6). The main
284 effect of spatial correlation is to channel flow along interconnected regions of high conductivity
285 (compare Figures 9-a and 9-b, see also Bruderer-Weng et al., 2004 for the effect of different
286 correlation lengths on flow channeling).

10
287
288 Flow patterns in anisotropic networks are shown in Figure 10 and 11 for 18 realizations with
289 different degrees of anisotropy (RP2)/(RT2), spatial correlation LP/LT, and tube size variability
290 COV(R2). The number of parallel tubes responsible for 50% of the total flow is included in
291 Figure 12 for all cases. Significant flow takes place along transverse tubes when transverse tubes
292 are much more conductive than parallel tubes (RP2)(RT2) (Figure 10-a), or when there is
293 high transverse correlation LP/LT1 (Figure 11-a - upper bound was labeled series-of-parallel
294 configuration in Figure 5). On the other hand, there is virtually no flow along transverse paths
295 when parallel tubes are much larger than the transverse tubes (RP2)(RT2) (Figure 10-c) or
296 when there is high longitudinal correlation LP/LT1 (Figure 11-c); in these cases, flow localizes
297 along linear flow paths and global conductivity is limited by the smallest tubes along their
298 longitudinal paths (third row in Figures 10 and 11 - referred to the parallel-of-series bound in
299 Figure 5). Therefore, the number of parallel tubes responsible for most of the flow decreases
300 with increasing COV(R2) in this case as well.
301
302 Lets define the network tortuosity factor as the ratio =(NCP/Nhom)2 between the total number of
303 tubes in the backbone of the critical path NCP and the number of tubes in a straight streamline
304 parallel to the global flow direction Nhom [details in David, 1993]. A critical path analysis in
305 terms of tube flow rate is used to compute the tortuosity factors for fluid flow (see Bernab and
306 Bruderer, 1998). Figure 10 and 11 show flow patterns and associated tortuosity values. In
307 agreement with visual patterns, there is a pronounced decrease in tortuosity when (RP2)>>
308 (RT2) in anisotropic uncorrelated fields (Figure 10) or when LP/LT>>1 in anisotropic correlated
309 fields with high COV(R2) (Figure 11).
310
311 Spatial correlation reduces the probability of small tubes being next to large ones, and leads to
312 more focused channeling of fluid flow through the porous network. This can be observed by
313 visual inspection of cases shown in Figure 11 in comparison to the corresponding ones in Figure
314 10.
315
316 Tube length variability. Simple geometrical analyses show that the distance between adjacent
317 pore centers is 2R for simple cubic packing and face-centered cubic packing, and 1.5 R for

11
318 tetrahedral packing, where R is the grain radius [see also Lindquist et al., 2000]. However, the
319 constant tube length assumption made in this study is only an idealization for real sediments
320 [Bryant et al., 1993]. For example, the distance between adjacent pore centers in Fontainebleau
321 and Berea sandstones ranges from 20 to 600m with most tube lengths between 130 and 200m
322 [Lindquist et al., 2000; Dong and Blunt, 2009].
323
324 While pore-to-pore distance varies in real sediments, we note that the hydraulic conductivity of
325 tubes is much more dependent on the radius than on the tube length (see Equation 4). Therefore,
326 the imposed variability in tube radius causes variability in tube conductivity q that could equally
327 capture tube length variability. Clearly, variations in tube length would imply a non-regular
328 network topology.
329
330 6. Conclusions
331 Grain size and formation history dependent pore size distribution and spatial variability
332 determine the hydraulic conductivity, immiscible fluid invasion and mixed fluid flow, resource
333 recovery, storativity, and the performance of remediation strategies. Numerical simulations with
334 porous networks permit the study of pore-size distribution, spatial correlation and anisotropy on
335 hydraulic conductivity and flow patterns in pervious media.
336
337 In most cases, the hydraulic conductivity decreases as the variance in pore size increases because
338 flow becomes gradually localized along fewer flow paths. As few as 10% of pores may be
339 responsible for 50% of the total flow in media with high pore-size variability. The equivalent
340 conductivity remains within Hashin and Shtrickman bounds.
341
342 Spatial correlation reduces the probability of small pores being next to large ones. There is more
343 focused channeling of fluid flow along interconnected regions of high conductivity and the
344 hydraulic conductivity is higher than in an uncorrelated medium with the same pore size
345 distribution.
346
347 The equivalent hydraulic conductivity in anisotropic correlated media increases as the correlation
348 length parallel to the flow direction increases relative to the transverse correlation. The hydraulic

12
349 conductivity in anisotropic uncorrelated pore networks is bounded by the two extreme parallel-
350 of-series and series-of-parallel tube configurations. Flow analysis shows a pronounced
351 decrease in tortuosity when pore size and spatial correlation in the flow direction are higher than
352 in the transverse direction.
353
354 While Poiseuille flow defines the governing role of pore size on hydraulic conductivity, the
355 numerical results presented in this manuscript show the combined effects of pore size
356 distribution and variance, spatial correlation, and anisotropy (either in mean pore size or in
357 correlation length). In particular, results show that the proper analysis of hydraulic conductivity
358 requires adequate interpretation of preferential flow paths or localization along interconnected
359 high conductivity paths, often prompted by variance and spatial correlation. The development of
360 flow localization will impact a wide range of flow related conditions including the performance
361 of seal layers and storativity, invasion and mixed fluid flow, contaminant migration and
362 remediation, efficiency in resource recovery, the formation of dissolution pipes in reactive
363 transport, and the evolution of fine migration and clogging.
364
365
366 Acknowledgments
367 Support for this research was provided by the National Science Foundation and by DOE/NETL
368 Methane Hydrate Project under contract DE-FC26-06NT42963.

13
369 7. References
370 Acharya, R.C., van der Zee, S.E.A.T.M., and Leijnse, A. (2004), Porosity-permeability properties
371 generated with a new 2-parameter 3D hydraulic pore-network model for consolidated and unconsolidated
372 porous media, Advances in water resources, 27, 707-723.
373
374 Albrecht, K.A., Logsdon, S.D., Parker, J.C., and Baker, J.C. (1985), Spatial variability of hydraulic
375 properties in the emporia series, Soil science society of America journal, 49, 1498-1502.
376
377 Al-Kharusi, A.S. and Blunt, M.J. (2007), Network extraction from sandstone and carbonate pore space
378 images, Journal of petroleum science and engineering, 56, 219-231.
379
380 Al-Kharusi, A.S. and Blunt, M.J. (2008), Multiphase flow predictions from carbonate pore space images
381 using extracted network models, Water resources research, 44, W06S01, doi:10.1029/2006WR005695.
382
383 Al-Raoush, R.I. and Wilson, C.S. (2005), Extraction of physically realistic pore network properties from
384 three-dimensional synchrotron X-ray microtomography images of unconsolidated porous media system,
385 Journal of hydrology, 300, 44-64.
386
387 Balhoff, M.T., Thompson, K.E., and Hjorts, M. (2007), Coupling pore-scale networks to continuum-
388 scale models of porous media, Computer & Geosciences, 33, 393-410.
389
390 Benson, C. (1993), Probability distributions for hydraulic conductivity of compacted soil liners, Journal
391 of geotechnical engineering, 119(3), 471-486.
392
393 Benson, C. and Daniel, D. (1994), Minimum thickness of compacted soil liners: II. Analysis and case
394 histories, Journal of geotechnical engineering, 120(1), 153-172.
395
396 Bernab, Y. and Bruderer, C. (1998), Effect of the variance of pore size distribution on the transport
397 properties of heterogeneous networks, Journal of geophysical research, 103, B1, 513-525.
398
399 Bjerg, P.L., Hinsby, K., Christensen, T.H., and Gravesen, P. (1992), Spatial variability of hydraulic
400 conductivity of an unconfined sandy aquifer determined by a mini slug test, Journal of hydrology, 135,
401 107-122.
402
403 Blunt, M.J. (2001), Flow in porous media pore-network models and multiphase flow, Current opinion in
404 colloid & interface science, 6, 197-207.
405
406 Bruderer, C. and Bernab, Y. (2001), Network modeling of dispersion: Transition from Taylor dispersion
407 in homogeneous networks to mechanical dispersion in very heterogeneous ones, Water resources research,
408 37(4), 897-908.
409
410 Bruderer-Weng, C., Cowie, P., Bernab, Y., and Main, I. (2004), Relating flow channelling to tracer
411 dispersion in heterogeneous networks, Advances in water resources, 27, 843-855.
412
413 Bryant, S.L., Mellor, D.W., and Cade, C.A. (1993), Physically representative network models of transport
414 in porous media, American institute of chemical engineers journal, 39(3), 387-396.
415
416 Cassel, D.K. (1983), Spatial and temporal variability of soil physical properties following tillage of
417 Norfolk loamy sand, Soil science society of America journal, 47, 196-201.
418

14
419 Dagan, G. (1979), Models of groundwater flow in statistically homogeneous porous formations, Water
420 resources research, 15(1), 47-63.
421
422 David, C. (1993), Geometry of flow paths for fluid transport in rocks, Journal of Geophysical Research,
423 98, 267-278.
424
425 Degroot, D.J. (1996), Analyzing spatial variability of in situ soil properties, Uncertainty in the geologic
426 environment, from theory to practice, Proceedings of uncertainty 96, Geotechnical special publication
427 No. 58, Madison, Wisconsin, 210-238.
428
429 Ditmars, J.D., Baecher, G.B., Edgar, D.E., and Dowding, C.H. (1988), Radioactive waste isolation in salt:
430 a method for evaluating the effectiveness of site characterization measurements, Final report ANL/EES-
431 TM-342, Argonne National Laboratory.
432
433 Dong, H. and Blunt, M. (2009), Pore-network extraction from micro-computerized-tomography images,
434 Physical review E, 80, 036307.
435
436 Duffera, M., White, J.G., and Weisz, R. (2007), Spatial variability of southeastern U.S. coastal plain soil
437 physical properties: implications for site-specific management, Geoderma, 137, 327-339.
438
439 Fatt, I. (1956)a, The network model of porous media, I. Capillary pressure characteristics, Trans. Am. Inst.
440 Min. Metall. Pet. Eng., 207, 144-159.
441
442 Fatt, I. (1956)b, The network model of porous media, II. Dynamic properties of a single size tube network.
443 Trans. Am. Inst. Min. Metall. Pet. Eng., 207, 160-163.
444
445 Fatt, I. (1956)c, The network model of porous media, III. Dynamic properties of networks with tube
446 radius distribution, Trans. Am. Inst. Min. Metall. Pet. Eng., 207, 164-177.
447
448 Fredd, C.N. and Fogler, H.S. (1998), Influence of transport and reaction on wormhole formation in
449 porous media, American Institute of Chemical Engineers Journal, 44(9), 1933-1949.
450
451 Freeze, R.A. (1975), A stochastic-conceptual analysis of one-dimensional groundwater flow in
452 nonuniform homogeneous media, Water resources research, 11(5), 725-741.
453
454 Griffiths, D.V. and Fenton, G.A. (1993), Seepage beneath water retaining structures founded on spatially
455 random soil, Gotechnique, 43(4), 577-587.
456
457 Griffiths, D.V. and Fenton, G.A. (1997), Three-dimensional seepage through spatially random soil,
458 Journal of geotechnical and geoenvironmental engineering, 123(2), 153-160.
459
460 Griffiths, D.V., Paice, G.M., and Fenton, G.A. (1994), Finite element modeling of seepage beneath a
461 sheet pile wall in spatially random soil, Computer methods and advances in geomechanics,
462 Siriwardane&Zaman, Balkema, Rotterdam, 1205-1209.
463
464 Gutjahr, A., Gelhar, L.W., Bakr, A.A., and MacMillan, J.R. (1978) Stochastic analysis of spatial
465 variability in subsurface flows 2. Evaluation and application, Water resources research, 14(5), 953-959.
466
467 Hashin, Z. and Shtrikman, S. (1962), A variational approach to the theory of the effective magnetic
468 permeability of multiphase materials, Journal of applied physics, 33(10), 3125-3131.
469

15
470 Hoefner, M.L. and Fogler, H.S. (1988), Pore evolution and channel formation during flow and reaction in
471 porous media, American Institute of Chemical Engineers Journal, 34(1), 45-54.
472
473 Hoeksema, R.J. and Kitanidis, P.K. (1985), Analysis of the spatial structure of properties of selected
474 aquifers, Water resources research, 21(4), 563-572.
475
476 Hoshen, J. and Kopelman, R. (1976) Percolation and cluster distribution. 1. Cluster multiple labeling
477 technique and critical concentration algorithm, Physical Review B, 14(8), 3438-3445.
478
479 Kang, Q., Tsimpanogiannis, I.N., Zhang, D., and Lichtner, P.C. (2005) Numerical modeling of pore-scale
480 phenomena during CO2 sequestration in oceanic sediments, Fuel processing technology, 86, 1647-1665.
481
482 Kampel, G, Goldsztein, G.H., and Santamarina, J.C. (2008), Plugging of porous media and filters:
483 Maximum clogged porosity, Applied physics letters, 92, 084101.
484
485 Lacasse, S., and Nadim, F. (1996), Uncertainties in characterizing soil properties, Proceedings of
486 uncertainty 96, Geotechnical special publication No. 58, Madison, Wisconsin, 49-75.
487
488 Landau, L. D., and Lifshitz, E. M. (1960), Electrodynamics of continuous media, Pergamon Press,
489 Oxford, New York.
490
491 Laudone, G.M., Matthews, G.P., and Gane, P.A.C. (2008), Modelling diffusion from simulated porous
492 structures, Chemical engineering science, 63, 1987-1996.
493
494 Libardi, P.L., Reichardt, K., Nielsen, D.R., and Biggar, J.W. (1980), Simple field methods for estimating
495 soil hydraulic conductivity, Soil science society of America journal, 44(1), 3-7.
496
497 Lindquist, W.B., Venkatarangan, A., Dunsmuir, J., Wong, T. (2000), Pore and throat size distributions
498 measured from synchrotron X-ray tomographic images of Fontainebleau sandstones, Journal of
499 geophysical research, 105, B9, 21509-21527.
500
501 Matheron, G. (1967), Elments Pour une Thorie des Milieux Poreux, Masson, Paris.
502
503 Mu, D., Liu, Z-S., Huang, C., and Djilali, N. (2008), Determination of the effective diffusion coefficient
504 in porous media including Knudsen effects, Microfluid Nanofluid, 4(3), 257-260.
505
506 Narsilio, G.A., Buzzi, O., Fityus, S., Yun, T.S., and Smith, D.W. (2009), Upscaling of Navier-Stokes
507 equations in porous media: theoretical, numerical and experimental approach, Computers and
508 Geotechnics, 36(7), 1200-1206.
509
510 Phadnis, H.S. and Santamarina, J.C. (2010), Pore size limited life in sediments. In preparation.
511
512 Prat, M. (2002), Recent advances in pore-scale models for drying of porous media, Chemical engineering
513 journal, 86, 153-164.
514
515 Reeves, P.C. and Celia, M.A. (1996), A functional relationship between capillary pressure, saturation, and
516 interfacial area as revealed by a pore-scale network model, Water resources research, 32(8), 2345-2358
517
518 Sahimi, M. (1994), Applications of percolation theory, Taylor&Francis, 258 pages.
519

16
520 Stauffer, D. and Aharony, A. (1992), Introduction to percolation theory, 2nd edition, Taylor&Francis, 181
521 pages.
522
523 Suicmez, V.S., Piri, M., and Blunt, M.J. (2008), Effects of wettability and pore-level displacement on
524 hydrocarbon trapping, Advances in water resources, 31, 503-512.
525
526 Surasani, V.K., Metzger, T., and Tsotas, E. (2008), Consideration of heat transfer in pore network
527 modelling of convective drying, International journal of heat and mass transfer, 51, 2506-2518.
528
529 Taskinen, A., Sirvio, H, and Bruen, M. (2008), Generation of two-dimensionally variable saturated
530 hydraulic conductivity fields: Model theory, verification and computer program, Computer &
531 Geosciences, 34, 876-890.
532
533 Tsimpanogiannis, I.N. and Lichtner, P.C. (2003), Pore network study of methane clathrate hydrate
534 dissociation. Eos, Trans, - Am. Geophys. Union 84(46) (Abstract OS51B-0848).
535
536 Tsimpanogiannis, I.N. and Lichtner, P.C. (2006), Pore-network study of methane hydrate dissociation,
537 Physical review E, 74, 056303.
538
539 Valvatne, P.H. (2004), Predictive pore-scale modeling of multiphase flow, PhD thesis, Department of
540 earth science and engineering, Imperial College.
541
542 Warren, J.E. and Price, H.S. (1961), Flow in heterogeneous porous media, Society of petroleum engineers,
543 153-169.
544
545 Warrick, A.W. and Nielsen, D.R. (1980), Spatial variability of soil physical properties in the field, In: Hill,
546 D. (Ed.), Applications of soil physics, Academic press, Inc., New York, 319-344.
547
548 Wiener, O. (1912), Abhandlungen der mathematisch. Physischen Klasse der Kniglichen Schsischen
549 Gesellschaft der Wissenschaften, 32, 509.

17
550 Table 1. Equivalent hydraulic conductivity Mixture models and Bounds
551
Equivalent keq Assumptions / Comments Reference

N N
Arithmetic mean - Parallel
k a = ti ki t i ti Parallel stratified media
i =1 i =1
ki Upper bound
Wiener
(1912)
Harmonic mean - Series
N N
ti
k h = ti ti ki Perpendicular stratified media (1D flow)
i =1 k
i =1 i
Lower bound

1N Geometric mean
N Warren and
k g = k i Lognormal ki distribution / isotropic media
1 Price (1961)
2D flow

Statistically homogeneous and isotropic Landau and

k eq = k a k h1
=
D 1
Lifshitz
D Weighted average of Wiener bounds (1960)

f 1 f 0 ( k1 k 0 ) 2 f 1 f 0 ( k1 k 0 ) 2
Based on a model constructed of composite Hashin and
ka k eq k a spheres. Isotropic binary medium. Uniform Shtrikman
k 0 ( D f 0 ) + k1 f 0 k1 ( D f 1 ) 2 + k 0 f 1
flow (1962)

if f 0 0.5 k eq k ac
if f 0 0.5 k eq k ac
if f 0 = 0.5 k eq = k1 k 0
where
k ac =
1
2
[
( f1 f 0 )(k1 k 0 ) + ( f1 f 0 ) 2 (k1 k 0 ) 2 + 4k1 k 0 ] Isotropic 2D random two phase mosaic
Matheron
if f 0 0.5 k eq k m medium.
(1967)
f 1 k 0 k1 + f 0 k a k 0 ( 2 k a k 0 )
Uniform flow.
km =
f1m * + f 0 k 0 ( 2 k a k 0 )
if f 0 0.5 k eq k m
f 0 k a + f1 k 0 ( 2m * k 0 )
k m = k 0 k1
f1k 0 k1 + f1m * k 0 ( 2 m * k 0 )

Note: keq = equivalent hydraulic conductivity; f0 and f1 are the fractions of the medium with hydraulic conductivity
k0 and k1, where k1> k0 , D=space dimension (i.e., 1, 2 or 3); m * = f 1 k 0 + f 0 k1

552

18
 1.0
0.8
(a) Fraction of small tubes = 20%

Probability
0.6
0.4
0.2
0.0
0.1 1 10 100 1000 10000
(R/[m])4

0.20
(b) (R2) = 1
Probability

0.15
COV(R2)=2.85
0.10
COV(R2)=0.49
0.05

0.00
0 1 2 3
(R/[m])2

COV=1
0.15
(c)
(R2) = 0.1 (R2) = 10 (R2) = 1000
Probability

0.10
(R2) = 0.1 (R2) = 10 (R2) = 1000
0.05

0.00
0.001 0.01 0.1 1 10 100 1000 10000

553 (R/[m])2
554
555
556 Figure 1. Schematic of typical distribution of R2 in spatially varying fields. (a) Bimodal
557 distribution of tubes (used in Figure 2 and 3) when fraction of small tubes is 20%. The relative
558 size of large to small tube radii is (RL/RS)4=103. (b) Two distributions of tube size R2 with the
559 same (R2) but different standard deviation used in Figure 4, 6, and 7. As the coefficient of
560 variation increases, the distribution of R2 is skewed to the right. (c) Distributions of R2 used in
561 Figure 5. Note that (R2) of two sets of tube size distribution with different (R2) are adjusted to
562 have same COV(R2).

19
563
1.000
RL/RS = 5.62
Normalized hydraulic conductivity kmix/kL (RL/RS)4 = 103
2D cn=4

2D cn=4
0.100 2D cn=6 2D cn=8

2D cn=6 3D cn=6

0.010
2D cn=8

0.001
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

Fraction of small tubes

564
565
566
567 Figure 2. Effect of coordination number cn on equivalent hydraulic conductivity in bimodal
568 distribution kmix normalized by the hydraulic conductivity in the field composed of only large
569 tubes kL. Each point is the average value of 20 realizations. Bimodal distribution of tubes. The
570 relative size of large L to small S tube radii is (RL/RS)4 = 103. Two-dimensional network model:
571 50x50 nodes, 4900 tubes, and cn=4 (circle) / cn=6 (triangle) / cn=8 (square). Three-dimensional
572 network model: 15x15x15 nodes, 9450 tubes, and cn=6 (diamond).

20
 1.000
RL/RS = 5.62
Wiener
upper - (RL/RS)4 = 103
Paralle
Normalized hydraulic conductivity kmix/kL

l
Hashi
n and S
htrikm
a n (2D
)
0.100

Matheron
Landa
u and
Lifshit
z
0.010
Ha s h Matheron Ge
in an om
d Shtri etr
kman ic
(2 D) me
an
Wiener
lower -
Series

0.001
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

Fraction of small tubes

573
574
575
576 Figure 3. Computed equivalent hydraulic conductivity in bimodal distribution kmix normalized
577 by the hydraulic conductivity in the field composed of only large tubes kL, models and bounds as
578 a function of the fraction of small tubes. Points represent the maximum (square), average
579 (triangle), and minimum (circle) values of 20 realizations at each fraction of small tubes. Bounds
580 and models are described in Table 1. Two-dimensional network model: 50x50 nodes, 4900 tubes,
581 bimodal distribution of tubes, relative size of large L to small S tube radii (RL/RS)4 = 103, and
582 coordination number cn=4.

21
583
Normalized hydraulic conductivity kdist/kmono 10.0

ka

1.0

keq
2D cn=8
kg
2D cn=6
3D cn=6
kh
2D cn=4
0.1
0.0 0.5 1.0 1.5 2.0

584 Coefficient of variation of R2

585
586
587 Figure 4. Equivalent hydraulic conductivity in uncorrelated tube network kdist normalized by the
588 hydraulic conductivity for the mono-sized tube network kmono as a function of the coefficient of
589 variation of R2. Each point is a single realization. All realizations have the same (R2). Two-
590 dimensional network model: 50x50 nodes, 4900 tubes, and cn=4 (empty triangle), cn=6 (empty
591 square), cn=8 (empty circle). Three-dimensional network model: 15x15x15 nodes, 9450 tubes,
592 and cn =6 (solid diamond). Shaded areas show arithmetic ka, geometric kg, harmonic kh mean of
593 2D and 3D system, and analytical solution keq of 3D system (Equation 3).

22
 10
Normalized hydraulic conductivity kdist/kmono

series-of-parallel
(RP2)/(RT2)
10-2
1
10-1

100
0.1

parallel-of-series 101

102
0.01
0.0 0.5 1.0 1.5 2.0

594 Coefficient of variation of RP2 or RT2

595
596
597 Figure 5. Anisotropic conductivity: Equivalent hydraulic conductivity in uncorrelated and
598 distributed tube network kdist normalized by the hydraulic conductivity for the mono-sized tube
599 network kmono as a function of coefficient of variation of R2. Normalized hydraulic conductivities
600 are obtained at different values of the ratio between the mean tube size parallel and transverse to
601 the flow direction (RP/RT)2. Each point is the average value of 20 realizations (using same set of
602 tube sizes, but different spatial distribution) . For clarity, results for intermediate sequences are
603 shown as shaded area. Normalized hydraulic conductivities in series of parallel and parallel of
604 series circuits are also obtained. Two-dimensional network model: 50x50 nodes, 4900 tubes,
605 cn=4, and log-normal distribution of R2.

23
 Normalized hydraulic conductivity kcor/kmono
1.0
(a)

L/D
30/39
15/39
5/39
uncorrelated
0.1
0.0 0.5 1.0 1.5 2.0 2.5

Coefficient of variation of R2
normalized hydraulic conductivity kcor/kmono

1
(b) L/D
30/39
Coefficient of variation of

0.8
15/39
0.6

5/39
0.4

0.2

uncorrelated
0
0.0 0.5 1.0 1.5 2.0 2.5

606 Coefficient of variation of R2

607
608
609 Figure 6. Correlated field. (a) Equivalent hydraulic conductivity in isotropic uncorrelated and
610 correlated tube network kcor normalized by the hydraulic conductivity for the mono-sized tube
611 network kmono as a function of the coefficient of variation of R2. (b) Coefficient of variation of the
612 equivalent hydraulic conductivities as a function of the coefficient of variation of R2. The
613 correlation length L is reported relative to the specimen size. Each point stands for the average of
614 100 realizations. Two-dimensional network model: 40x40 nodes, 3120 tubes, cn=4, and log-
615 normal distribution of R2.

24
 10
Normalized hydraulic conductivity kCOV>0 / kmono

1
COV=0.5

0.1
COV=1.4

0.01 COV=2.8

0.001
0.01 0.1 1 10 100
LP/LT

LP/LT
LP 2/39 2/39 30/39
LT LP
LT 30/39 2/39 2/39

616
617
618
619 Figure 7. Effect of anisotropic correlation on equivalent hydraulic conductivity in an
620 anisotropically correlated tube network kCOV>0 normalized by the hydraulic conductivity for the
621 mono-sized tube network kmono. Three sets of tubes different COV(R2) are generated and used to
622 form correlated fields of different anisotropic correlation length. LP and LT are the correlation
623 lengths parallel and transverse to flow direction. D is the length of medium perpendicular to the
624 flow direction. In the range between LP/LT=0.01 to 1, LP=2D/39 fixed and LT changes from
625 2D/39 to 30D/39. In the range between LP/LT=1 to 100, LT=2D/39 fixed and LP changes from
626 2D/39 to 30D/39. Each point is an average of 20 realizations. Two-dimensional network model:
627 40x40 nodes, 3120 tubes, cn=4, and log-normal distribution of R2.

25
 Fraction of small tubes
(a) 20% 50% 80%

0.5

0.4
tubes50%

0.3 total-tube50% small-tube50%

0.2
large-tube50%
0.1
(b)
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

628 Fraction of small tubes

629
630
631 Figure 8. Analysis of flow pattern in network model of bimodal distribution of R2 (2D cn=4 -
632 Percolation occurs when the fraction of small tubes is 0.5. Refer to Figure 2 and 3 for simulation
633 details). (a) Flow intensity in each tube of the network of different fraction of small tubes. The
634 change of flow pattern in each fraction of small tubes is well detected. The arrow indicates the
635 predominant fluid flow direction. (b) Fraction of tubes tube50% responsible for 50% of total
636 conductivity. The fraction total-tube50% is the summation of small-tube50% and large-tube50%.

26
 COV(R2)=0.49 COV(R2)=1.26 COV(R2)=1.95

(a) Uncorrelated network

(Fig. 4 and 6)
(b) Isotropically correlated
Network (Fig. 6)

0.5

0.4
tubes50%

0.3

0.2
uncorrelated
0.1 correlated
(c)
0.0
0.0 0.5 1.0 1.5 2.0 2.5

637 Coefficient of variation of R2

638
639
640 Figure 9. Analysis of flow pattern in network model of log-normal distribution of R2 (refer to
641 Figure 4 and 6 for simulation details). (a) Flow intensity in each tube in spatially uncorrelated
642 network. (b) Flow intensity in each tube in spatially correlated networks. Thickness of line
643 represents the intensity of flow rate. (c) Fraction of tubes tube50% responsible for 50% of total
644 conductivity.

27
 COV(R2)=0.33 COV(R2)=1.0 COV(R2)=1.6
(a) (RP2)/ (RT2)=0.01

= 375.9 = 243.7 = 179.8

(b) (RP2)/ (RT2)=1.0

= 1.1 = 2.6 = 4.2

(b) (RP2)/ (RT2)=100

645 = 1.0 = 1.0 = 1.0

646
647
648 Figure 10. Analysis of flow patterns in anisotropic networks made of tubes with the same mean
649 size (R2) but different variance in size as captured in COV(R2) (refer to Figure 5 for simulation
650 details). Anisotropy ratios: (a) (RP2)/(RT2)=0.01, (b) (RP2)/(RT2)=1.0, (c) (RP2)/(RT2)=100.
651 The arrow indicates the global flow direction. The line thickness used to represent the tubes is
652 proportional to the flow intensity in each tube. Tortuosity values are shown for each case.

28
 COV(R2)=0.5 COV(R2)=1.4 COV(R2)=2.8
(a) LP/LT=1/15

= 1.6 = 60.0 = 30.1

(b) LP/LT=1/1

= 8.0 = 7.1 = 6.6

(c) LP/LT=15 /1

653 = 1.9 = 1.7 = 2.1

654
655
656 Figure 11. Analysis of flow pattern in anisotropically correlated networks made of three sets of
657 tube areas with different COV(R2) (refer to Figure 7 for simulation details). Anisotropy ratios:
658 (a) LP/LT=1/15, (b) LP/LT=1/1, and (c) LP/LT=15/1. The arrow indicates the global flow direction.
659 The line thickness used to represent the tubes is proportional to the flow intensity in each tube.
660 Tortuosity values are shown for each case.

29
 0.5 0.5

0.4 (RP2)/(RT2) 0.4

tubes50%
tubes50%

0.3 102 0.3

0.2 100 0.2 COV(R2)=0.5

0.1 0.1
10-2 COV(R2)=1.4
(a) (b) COV(R2)=2.8
0.0 0
0 0.5 1 1.5 2 0.01 0.1 1 10 100

Coefficient of variation of R2 LP/LT

661
662
663 Figure 12. Fraction of tubes tubes50% carrying 50% of the total flux as a function of (a)
664 coefficient of variation of R2 in anisotropically uncorrelated field and (b) the ratio of parallel to
665 transverse correlation length LP/LT.

30