Advances in Water Resources 96 (2016) 373–388

Contents lists available at ScienceDirect

Advances in Water Resources

journal homepage: www.elsevier.com/locate/advwatres

Influence of surface roughness on nonlinear flow behaviors in 3D

self-affine rough fractures: Lattice Boltzmann simulations
Min Wang a,b, Yi-Feng Chen a,∗, Guo-Wei Ma b, Jia-Qing Zhou a, Chuang-Bing Zhou a,c
a
State Key Laboratory of Water Resources and Hydropower Engineering Science, Wuhan University, Wuhan, 430072, China
b
School of Civil, Environmental and Mining Engineering, University of Western Australia, Perth, 6009, Australia
c
School of Civil Engineering and Architecture, Nanchang University, Nanchang, 330031, China

a r t i c l e i n f o a b s t r a c t

Article history: This study investigates the impacts of surface roughness on the nonlinear fluid flow through three-
Received 22 May 2016 dimensional (3D) self-affine rock fractures, whose original surface roughness is decomposed into pri-
Revised 2 August 2016
mary roughness (i.e. the large-scale waviness of the fracture morphology) and secondary roughness (i.e.
Accepted 12 August 2016
the small-scale unevenness) with a wavelet analysis technique. A 3D Lattice Boltzmann method (LBM)
Available online 13 August 2016
is adopted to predict the flow physics in rock fractures numerically created with and without consider-
Keywords: ation of the secondary roughness, respectively. The simulation results show that the primary roughness
Self-affine fractures mostly controls the pressure distribution and fracture flow paths at a large scale, whereas the secondary
Surface roughness roughness determines the nonlinear properties of the fluid flow at a local scale. As the pressure gradient
Roughness decomposition increases, the secondary roughness enhances the local complexity of velocity distribution by generat-
Nonlinear fluid flow ing and expanding the eddy flow and back flow regions in the vicinity of asperities. It was found that
Lattice Boltzmann method
the Forchheimer’s law characterizes well the nonlinear flow behavior in fractures of varying roughness.
The inertial effects induced by the primary roughness differ only marginally in fractures with the rough-
ness exponent varying from 0.5 to 0.8, and it is the secondary roughness that significantly enhances the
nonlinear flow and leads to earlier onset of nonlinearity. Further examined were the effects of surface
roughness on the transmissivity, hydraulic aperture and the tortuosity of flow paths, demonstrating again
the dominant role of the secondary roughness, especially for the apparent transmissivity and the equiva-
lent hydraulic aperture at high pressure gradient or high Reynolds number. The results may enhance our
understanding of the role of surface roughness in the nonlinear flow behaviors in natural rock fractures.
© 2016 Elsevier Ltd. All rights reserved.

1. Introduction vides an alternative, powerful tool for elaborating the flow process
provided that the rough geometries of fracture surfaces are suffi-
Characterizing the fluid flow process through individual rough ciently represented in the numerical models.
fractures is crucial to the understanding of mass and energy trans- Macroscopically, the fluid flow in a single rock fracture is com-
port behaviors in fractured rocks, aquifers and reservoirs related to monly approximated by the cubic law (Witherspoon et al., 1980),
water resource utilization, enhanced oil recovery, geothermal en- which is derived from the analogy of laminar flow between two
ergy development, remediation of non-aqueous phase liquids, nu- perfectly smooth parallel plates separated from each other by a
clear waste disposal and carbon capture and storage (Berkowitz, constant gap (denoted by aperture e). For rough-walled fractures
20 02; Neuman, 20 05; Sahimi, 2011). Laboratory- and field-scale with variable apertures, the cubic law has been expanded by re-
measurements are of great importance for examining and quan- placing the aperture e of the idealized parallel smooth fracture
tifying this flow process, but may prove insufficient because most with the hydraulic aperture eh . For more accurately representing
experimental techniques fail to directly observe the details of the the fluid flow through fractures, numerical methods were used,
flow behaviors in real rock fractures locally influenced by the com- among which the Reynolds lubrication equation was introduced
plex surface geometry at varying scales, such as channeling flow, as an alternative approximation to the considerably more compu-
eddy flow, back flow and turbulent flow. Numerical modeling pro- tationally intensive solution of the Navier-Stokes (NS) equations.
This approximation has been widely used to estimate the hydraulic
conductivities of rough-walled fractures in the last two decades
∗
Corresponding author. Fax: +86-27-68774295.
(Zimmerman and Bodvarsson, 1996; Brush and Thomson, 2003;
E-mail addresses: [email protected] (M. Wang), [email protected] (Y.-F.
Chen), [email protected] (G.-W. Ma).
Lee et al., 2014; Liu et al., 2016b), but fails to simulate the inertial

http://dx.doi.org/10.1016/j.advwatres.2016.08.006
0309-1708/© 2016 Elsevier Ltd. All rights reserved.
374 M. Wang et al. / Advances in Water Resources 96 (2016) 373–388

Table 1
A review on the roughness parameters for fracture profiles/surfaces.

Parameter Description Method Expression/Relationship

JRC Joint roughness coefficient Empirical method: dimensionless parameter 1<JRC<20

ξ Peak asperity height Statistical method: asperity parameter ξ = ξmax − ξmin

n
Za Mean asperity height line/plane Statistical method: asperity parameter Za = 1n ξi

i=1

n
RMS Root-mean-square asperity height value Statistical method: asperity parameter RMS = 1
n
(ξi − ξa )2
i=1

n
Rm Mean asperity height Statistical method: asperity parameter Rm = 1
n
|ξi − ξa |
i=1

n
Rsk Skewness: average of the first derivative of the Statistical method: asperity parameter Rsk = 1
nRMS3
(ξi − ξa )3
i=1
surface; negative for valleys; positive for
peaks

n
Rku Kurtosis: a measure of sharpness of surface Statistical method: asperity parameter Rku = 1
nRMS4
(ξi − ξa )4
i=1
peaks


n
Z1 Root-mean-square value of asperity height Statistical method: asperity parameter Z1 = (n−11)x ξi2

i=1

n
Z2 Root-mean-square of first derivative of asperity Statistical method: asperity parameter Z2 = 1
(n−1 )(x )2
(ξi+1 − ξi )2
i=1
height


n
Z3 Root-mean-square of second derivative of Statistical method: asperity parameter Z3 = 1
(n−1 )(x ) 3 (ξi+1 − 2ξi + ξi−1 )2
i=1
asperity height

n 
m
xip − xnj
i=1
Characterize the roughness anisotropy, Z4 = 0 Z4 = , -1<Z4 <1
j=1
Z4 Statistical method: asperity parameter n m
xip + xnj
for isotropy i=1 j=1

n
ξi+1 −ξi
θP Mean angles of the inclinations of the profile Statistical method: angularity parameter θP = tan−1 ( n−1
1
| x |)
i=1
At
Rs Surface roughness coefficient Statistical method: surface parameter Rs = An
, Rs ≥1
Ts Surface tortuosity coefficient Statistical method: surface parameter Ts = At
An
cos ϕ ,0<Ts ≤Rs

n
S(d) Structure function: mean square of the Random field method S (d ) = 1n (ξi+d − ξi )2
i=1
difference in height expected over any
spatial distance d

n
C(d) Auto correlation function: spatial variations of Random field method C (d ) = lim 1
(ξi − ξa )(ξi+d − ξa )
n→∞ n
i=1
roughness parameter
H Hurst exponent Fractal method: dimensionless parameter 0<H<1
Df Fractal dimension Fractal method: dimensionless parameter D f =DE -H

Note: n is the number of data points, ξ i is the asperity height at point i, ξ max is the maximum asperity height, ξ min is the minimum asperity height, x is the
distance on the axial direction between point i+1 and i, xi p is the distance along the profile that the slop is positive, xj n is the distance along the profile
that the slop is negative, d is arbitrary spatial distance, DE is the Euclidean dimension (1 for the line profiles and 2 for the surface profiles), and ϕ is the angle
between the normal of the surface formed by the four extreme points and the normal of the mean plane of the joint surface.

flows that may occur in many fractured aquifers. To characterize been increasingly applied for the simulation of complex flow in
the nonlinear flow through fractures, the complete NS equations fractured aquifers (Kim et al., 2003; Madadi et al., 2003; Briggs
including the acceleration and inertial terms should be adopted et al., 2014), ranging from the prediction of permeability to the
as the governing equation. But unfortunately, the existence and study of wettability effect in 2D or 3D self-affine rough fractures
uniqueness of closed-form solution of the full NS equations in 3D (Gutfraind and Hansen, 1995; Madadi and Sahimi, 2003; Eker and
fractures is not proven yet (Zimmerman and Bodvarsson, 1996; Akin, 2006; Dou et al., 2013). The application of LBM is, however,
Wang et al., 2015). rarely reported for simulations of the nonlinear flow in 3D rough
Over the past decades, various Computational Fluid Dynam- fractures.
ics models have been developed to solve the highly nonlinear NS The heterogeneity or roughness of fracture surfaces is of impor-
equations for flow in different branches of geosciences. Among tance in qualifying the flow and transport processes through frac-
them, the Lattice Boltzmann method was proposed as a discrete tured rocks, especially when the nonlinear flow occurs as a result
approximation of the incompressible NS equations based on ki- of non-negligible inertial losses from variations in flow velocity or
netic theory rather than continuum theory (Wolf-Gladrow, 20 0 0). direction along the flow paths due to local changes of geometries
This method has evolved as a strong tool to simulate fluid flow in (Chen et al., 2015). The surface roughness is commonly character-
complex geometries (Succi et al., 1991; Llewellin, 2010), because ized with four main kinds of roughness parameters (Barton and
the no slip boundary (on the solid/liquid interface) can be im- Choubey, 1977; Belem et al., 20 0 0; Schmittbuhl et al., 20 08; Gao
plemented flexibly (Chen et al., 1996; Filippova and Hänel, 1998, and Wong, 2015), as summarized in Table 1. It was defined by
Mei et al., 20 0 0; Mohammadipoor et al., 2014). A great quan- the ISRM (International Society of Rock Mechanics) at two concep-
tity of extensions to LBM has been developed and validated to tual scales (ISRM, 1978): the large-scale waviness and the small-
simulate the single-phase or multi-phase flow from microscopic scale unevenness. With this conceptual understanding, Jing et al.
to macroscopic scale (Chen and Doolen, 1998; Dardis and Mc- (1992) proposed a concept of primary and secondary roughness to
Closkey, 1998; Kang et al., 2002; Anwar and Sukop, 2009; Chen explain the mechanical behaviors of rock fractures under shear-
et al., 2014; Saravanathiiban et al., 2014; Zhang et al., 2016), the ing, where the primary roughness (i.e. the large-scale waviness)
solute transport (Zhang et al., 2008; Ginzburg 2013), the flow in- was defined as the dominating and larger-scale wavy surface un-
teracting with deformable bodies (De Rosis 2014a, 2014b) and the dulation and the secondary roughness as the randomly distributed
nonlinear flow (Pan et al., 2006; Jeong et al., 2006; Sukop et al., small-scale unevenness superimposed on the primary waviness
2013; Newman and Yin, 2013) in porous media. LBM has also surfaces.
 M. Wang et al. / Advances in Water Resources 96 (2016) 373–388 375

2.4
4.0
2.2
3.6
2.0
3.2
1.8
2.8
H=0.50 2.4 H=0.60 1.6
0 0
1.4
2.0
5 5 1.2
1.6
10 10 1.0
0.8
15 15 0.8
20 0.4 20 0.4
25 Unit: mm 25
) ) 0.2
0 30 mm 0 30 mm Unit: mm
5 35 X( 5 35 X(
10 40 10 40
Y (m 15 Y (m 15
m) 20 45 m) 20 45
25 50 25 50
2.0 2.0
1.8 1.8
1.6 1.6
1.4 1.4
1.2
H=0.70 1.2 H=0.80
1.0 1.0
0 0
5 0.8 5 0.8
10 0.4 10 0.4
15 0.2 15 0.2
20 20
Unit: mm Unit: mm
25 25
) )
0 30 mm 0 30 mm
5 35 X( 5 35 X(
10 10
40 Y (m 15 40
Y (m 15
m) 20 45 m) 20 45
25 50 25 50

Fig. 1. Four self-affine surfaces generated by the modified SRA algorithm.

The concept of fracture roughness decomposition was followed variance of the increments, the statistical self-affine property of
by Zou et al. (2015) with a wavelet analysis technique, who found fBm can be expressed as:
by numerical simulations that the secondary roughness plays a sig-
nificant role in the nonlinear flow through a 2D rough fracture
Z (x + rhx , y + rhy ) − Z (x, y ) = 0 (1)
by forming a thin boundary layer with nonlinear dynamic phe-
nomenon of eddy flow over the whole domain. The 2D simula- σ 2 ( r ) = r 2H σ 2 ( 1 ) (2)
tions, however, suffer from the disadvantage for overestimating the
where  ·  represents the mathematical expectation, H is the
effect of roughness on reducing the effective advective flow area,
roughness exponent or Hurst exponent varying from 0 to 1 and re-
and hence fail to capture the flow behaviors through the most con-
lated to the 3D fractal dimension (Df ) by Df = 3 - H, r is a constant,
ductive paths (i.e., high aperture regions) and around the low aper-
ture regions. In this study, we conducted the numerical simulations and σ 2 is the variance defined as a function of r:
 
of single phase flow through eight 3D self-affine fracture models, σ 2 (r ) = [Z (x + rhx , y + rhy ) − Z (x, y )]2 (3)
four of which were generated from their counterparts by rough-
ness decomposition using wavelet analysis. A LBM-based code with Based on the scaling properties given by Eqs. (1) and (2), a
dimensional mapping between the physical units and LBM sim- number of SRA algorithms have been developed (e.g. Liu et at.,
ulation units was developed to simulate the fluid flow in rough 2004; Ye et al., 2015) for self-affine surface generation, among
fractures by recovering the solution of the full NS equations with- which Liu’s algorithm (Liu et at., 2004) improves the poor scal-
out loss of the inertial term. The effects of the surface roughness ing and correlation properties of fractal distributions involved in
with different frequencies on the fluid flow behaviors were ana- the traditional SRA algorithms. By using Liu’s algorithm, four self-
lyzed and discussed. Finally, the effects of surface roughness on the affine rough fracture wall surfaces of 50 mm in length and 25 mm
transmissivity, hydraulic aperture and the tortuosity of flow paths in width were generated, with H = 0.5, 0.6, 0.7 and 0.8, respec-
were further examined. tively, as shown in Fig. 1. Obviously, a higher value of H results in
a higher spatial correlation and a smoother surface, and vice versa.
It is to be noted that the Hurst exponent of natural rock frac-
2. 3D self-affine fractures and roughness decomposition tures has proved to vary mostly between 0.45 and 0.85 (corre-
sponding to a 3D fractal dimension between 2.55 and 2.15) (Odling,
2.1. Self-affinity of rough fracture surfaces 1994; Schmittbuhl et al., 2008; Babadagli et al., 2015). Boffa et al.
(1998) used high-resolution roughness measurements on fractured
As a matter of fact that the surface roughness of natural frac- granite and sandstone samples, and found that the rock surfaces
tures typically follows a self-affine fractal distribution (Mandelbrot, display self-affine geometries with a characteristic Hurst exponent
1983; Brown and Scholz, 1985; Odling, 1994), the geometry of nat- H = 0.8 ± 0.5 and H = 0.47 ± 0.5, respectively. Develi and Babadagli
ural rough fracture surfaces is commonly simulated with a frac- (1998) applied variogram analysis (VA) to evaluate the self-affinity
tional Brownian motion (fBm). The most widely-used methods are of natural crystallized limestone fracture surfaces, showing that the
the successive random additions (SRA), the randomization of the Hurst exponent ranges between 0.425 and 0.795. In this study, we
Weierstrass-Mandelbrot function and the Fourier transformation. measured the surface morphology of a granite fracture and a mar-
In this study, the SRA algorithm is adopted to generate rough frac- ble one with a high-resolution noncontact 3D Laser Scanner (Fig. 2)
ture wall surfaces for its efficiency and straightforwardness. (Li et al., 2014), and examined the Hurt exponent values of each
More specifically, we represent the surface height of a fracture profile consisting of about 500 data points with the VA method. A
wall with a continuous and single-valued function Z(x, y). In the total of 200 profiles were evaluated, and the calculated Hurt expo-
fBm framework, the stationary increment [Z(x+hx , y+hy )-Z(x, y)] nent values range between 0.52 and 0.87. Fig. 3 shows an analysis

over the distance (lag) h = h2x + h2y displays a Gaussian distribu- of the scaling properties related to the granite and marble frac-
tion with mean zero (Molz et al., 1997). In terms of the mean and ture surfaces over the entire topographical data. The relationship
376 M. Wang et al. / Advances in Water Resources 96 (2016) 373–388

Fig. 2. Measurement and characterization of self-affine rough fracture surfaces.

2.2. Roughness decomposition with wavelet analysis

The roughness of fracture surfaces is often represented by lo-

cal changes of geometries at different scales, which have dif-
ferent effects on the hydraulic behaviors of rock fractures (Zou
et al., 2015). The surface may be interpreted as a combination
of waviness and unevenness/roughness, as suggested by the ISRM
(1978). A mathematical tool for quantitative and objectively unique
roughness decomposition is required for realistic investigations of
hydro-mechanical and transport processes in natural rock frac-
tures. The methods of image analysis (Zawada-Tomkiewicz, 2010),
Fourier transform (Chae et al., 2004), mathematical morphology
(Młynarczuk, 2010) and wavelet analysis (Chamoli et al., 2007)
have been used to distinguish and filter the roughness components
in 2D fracture profiles or 3D fracture surfaces. Among them, the
wavelet analysis method has a pronounced advantage for the anal-
ysis of the multi-scale effects of surface roughness on the fluid
flow through rock fractures, by means of extracting the differ-
ent frequencies of surface topographic information at different fre-
quency ranges (levels) and scales (Chae et al., 2004; Khoshelham
et al., 2011). Compared to the Fourier transform, in which signals
are represented as a sum of stationary sinusoids, the wavelet anal-
ysis is more suitable for analyzing the geometry of natural rock
fracture surfaces composed of non-stationary primary and station-
ary secondary roughness components (Zou et al., 2015). Conse-
quently, the effects of the different frequencies of surface topo-
graphic information on flow velocity fields in rock fractures can be
investigated numerically in combination or independently.
A fundamental issue associated with the fracture roughness de-
composition is to properly define a cut-off limit for separating
roughness from waviness, which ensures that the dominating fea-
tures of the undulation/waviness are maintained and meanwhile
the secondary roughness could be treated as a Gaussian white
noise process. In this study, the quantitative criterion proposed by
Zou at al. (2015) using the wavelet analysis method was followed.
The Db8 wavelet of the Daubechies wavelet family available in
the wavelet toolbox in MATLAB was invoked for its more suitabil-
ity for surface roughness investigation (Mehrishal and Sharifzadeh,
Fig. 3. Log-log plot of the standard deviation σ with the lag r for (a) a granite 2013). The discrete wavelet transform process, known as the Mal-
fracture surface and (b) a marble fracture surface.
lat’s pyramidal algorithm (Mallat, 1989), was applied for roughness
decomposition in a level-by-level procedure.
As an example, Fig. 4 shows the surface decomposed and re-
constructed by eight levels of decomposition, denoted by A1–
A8 and D1–D8, respectively, for the fracture surface with H = 0.6
between the variance and lag illustrates that the best-fitted values
(see Fig. 1). The approximation surfaces Ai at decomposition level
of H for the granite and marble surfaces are rather close to each
i mainly represents the dominant geometrical feature of non-
other, with H = 0.786 and 0.734, respectively.
 M. Wang et al. / Advances in Water Resources 96 (2016) 373–388 377

Approximation Surfaces Detail Surfaces

A1 D1
50 50
2 45 45
40 0.05 40
35 35
0 30 -0.05 30
25 25 25 25
20 20 X(mm) 20 20 X(mm)
15 15 15 15
Y (m 10 10 Y(m 10 10
m) 5 5 m) 5 5
0 0 00

A2 D2
50 50
45 0.05 45
2 40 40
35 35
0 30 -0.05 30
25 25 25 25
20 20 X(mm) 20 20 m)
15 15 15 15 X(m
Y(m 10 10 Y(m 10 10
m) 5 5 m) 5 5
00 00

A3 D3
50 50
45 45
2 40 0.05 40
35 35
0 30 -0.05 30
25 25 25 25 )
20 20 m) 20 20
X(m
m
15 15 X(m 15 15
Y(m 10 5 5
10 Y(m 10 5 5
10
m) m)
00 00

A4 D4
50 50
45 0.05 45
2 40 40
35 35
0 30 -0.05 30
25 25 25 25
20 20 m) 20 20 X(mm)
15 15 X(m 15 15
Y(mm 10 10 Y(m 10 10
) 5 5 m) 5 5
00 00

A5 D5
50 50
2 45 45
40 0.05 40
35 35
0 30 -0.05 30
25 25 25 25
20 20 m) 20 20 X(mm)
15 15 X (m 15 15
Y(m 10 10 Y(m 10 10
m) 5 5 m) 5 5
00 00

A6 D6
50 50
2 45 0.10 45
40 40
35 35
0 30 -0.10 30
25 25 25 25
20 20 X(mm ) 20 20 X(mm)
15 15 15 15
Y(m 10 10 Y(m 10 10
m) 5 5 m) 5 5
00 00

A7 D7
50 50
45 0.10 45
2 40 40
35 35
0 30 -0.10 30
25 25 25 25
20 20 X(mm ) 20 20 m)
15 15 15 15 X(m
Y(m 10 10 Y(m 10 5 5
10
m) 5 5 m)
00 00

A8 D8
50 50
45 45
2 40 0.10 40
35 35
0 30 -0.10 30
25 25 ) 25 25
20 20
X (m
m 20 20 X(mm)
15 15 15 15
Y(m 10 5 5
10 Y(m 10 10
m) m) 5 5
00 00

0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 (mm) -0.10 -0.08 -0.06 -0.05 -0.03 -0.01 0.01 0.03 0.05 0.06 0.08 0.10 (mm)

Fig. 4. Eight levels wavelet decomposition of the roughness for the fracture surface with H = 0.6.
378 M. Wang et al. / Advances in Water Resources 96 (2016) 373–388

stationarity (i.e. the primary undulation/waviness), and the detail nent can be extracted by taking advantage of the central properties
surfaces Di mainly represents the small-scale unevenness or white of the self-affine function (Eq. (2)). This means that both the sec-
noise with higher frequency and smaller wave length (i.e. the sta- ondary and primary roughness of the asperities can be well quali-
tionary secondary roughness). Fig. 5 shows that the optimal cut-off fied with the Hurst/roughness exponent.
level of the surface appears at level 5 where the variance of the
primary roughness starts to change remarkably while the distri- 2.3. Aperture distribution
bution of the asperities of the secondary roughness is proven to
follow a Gaussian distribution by the Shapiro-Wilk’s (1965) test. It has been revealed by experimental measurements (Brown
Therefore, the approximation surface at level 5 was taken as the et al., 1998; Scesi and Gattinon, 2007; Indraratna et al., 2014)
primary roughness, and the sum of the detail surfaces at levels 1– and numerical simulations (Zimmerman and Bodvarsson, 1996;
5 was used to create the secondary roughness. Auradou, 2009; Xie et al., 2015) that the flow and transport of
It has been understood that the Hurst/roughness exponent of fluid through a fractured medium is highly sensitive to fracture
self-affinity is constant over a large range of scales and positions apertures. In previous studies, Gaussian distribution (Montemagno
(Odling, 1994), and hence the wavelet-based techniques could be and Pyrak-Nolte, 1999; Walsh et al., 2008) was mostly assumed to
used for Hurst exponent measurements (Simonsen et al., 1998; characterize the aperture distributions for both natural and artifi-
Chamoli et al., 2007). By means of changing the scale of the cial fractures. In this study, the shear displacement model devel-
wavelet, one is able to derive a scaling relationship between the oped by Wang et al. (1988) was applied to obtain the apertures
wavelet amplitudes at different scales, from which the Hurst expo- with Gaussian distribution based on the self-affine fractal model,
in which the aperture distribution b is simply generated by

Z (x + xd , y + yd ) − Z (x, y ) + b0 if Z (x + xd , y + yd ) > Z (x, y ) − b0
b(x, y ) = (4)
0 if Z (x + xd , y + yd ) ≤ Z (x, y ) − b0

where d = (xd , yd ) represents the shear displacement and b0 is

the mean separation between the lower and upper surfaces. When
d = (0, 0), the local apertures of the fracture become a constant
(b0 ). In a more general case, the two surfaces do not match be-
cause of the displacement relative to each other. Based on the self-
affinity of fBm (Eqs. (1)-(3)) and the aperture defined in Eq. (4), the
non-zero aperture distribution can be easily described by a Gaus-
sian distribution with a mean of b0 and a standard deviation of σ :

1 (b−b0 )2
f (b ) = √ e− 2σ 2 f or b ≥ 0 (5)
2πσ
As an example, Fig. 6(a) shows a stochastic realization of aper-
ture that follows the Gaussian distribution with b0 = 0.50 mm,
σ = 0.15 mm and d = (0, 31) for the generated surface with H = 0.6
(see Fig. 1). Fig. 6(b) indicates that the calculated aperture dis-
tribution from the generated fracture fits well with the theoreti-
cal Gaussian distribution function. Note that although the 3D frac-
ture surface and the aperture distribution generated by the fractal
model are artificial, it appears as a relatively more realistic concep-
tual presentation of natural rough fractures than the 2D fracture
model constructed by profiles.

3. Lattice Boltzmann simulation of fluid flow

3.1. Lattice Boltzmann method

LBM provides an alternative numerical method for simulating

hydrodynamic systems governed by the NS equations. The LBM
follows a ‘bottom-up’ approach (Benzi et al., 1992; Wolf-Gladrow,
20 0 0), where it starts from a discrete lattice model and has been
proven to recover the NS solution at macroscopic scale (Qian et al.,
1992; Succi 2001). In LBM, the fluid phase is discretized according
to the conservation of macroscopic properties and the necessary
symmetries required by the NS equations. A three-dimensional
D3Q19 lattice model was used in this study (Qian et al., 1992),
where D is the space dimension and Q is the number of discrete
velocity vectors in the model. The average movement of fluid par-
ticles is represented by a distribution function, fi (x), which is de-
fined for each lattice vector ei (the mesoscopic velocity in the ith
Fig. 5. (a) Variance of the reconstructed data against the decomposition level, with direction) at each site x. The mesoscopic velocities for the D3Q19
zero indicating the initial data before decomposition; (b) histogram of the sec- lattice model are illustrated in Fig. 7. The distribution function sat-
ondary roughness for the fracture surface with H = 0.6 (see Fig. 1). isfies the discrete Boltzmann equation at time t using the Lattice
 M. Wang et al. / Advances in Water Resources 96 (2016) 373–388 379

erns the rate at which the fluid tends towards equilibrium, with
τ = υ /cs 2 t + 0.5 in the lattice units. υ is the kinematic viscosity
of fluid
√ and cs is the lattice pseudo-sound-speed, which is equal
to 1/ 3. At time t, the equilibrium distribution function depends
on the fluid macroscopic density (ρ ) and velocity (u) at a node x,
which can be computed as:


18
ρ (x ) = f i (x ) (7)
i=0

1 
18
u= f i ( x )ei (8)
ρ (x ) i=0

According to the incompressible LB model presented by He and

Luo (1997), the equilibrium distribution function for model D3Q19
is:
9 3
f i ( x ) = ωi ρ ( x ) + ρ0 3 ( e i · u ) + ( e i · u ) − ( u · u )
eq
(9)
2 2
where ρ 0 is the mean density and ωi is the weight factor in the
ith direction. The weight factors for D3Q19 are:
1/3, i = 0,
ωi = 1/18, i = 1, 2, 3, 4, 5, 6,
1/36, i = 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18
(10)
In LBM, pressure is defined as:

P (x ) = cs2 ρ (x ) (11)

3.2. Boundary conditions and dimensional mapping

One of the most important and crucial issues in the LB sim-

ulation of flow is the accurate modeling of the boundary con-
ditions (Mohamad, 2011). Specifying the boundary conditions for
Fig. 6. (a) Realization of a self-affine fracture with the upper surface displaced from
the NS equations is somehow straightforward, but this is not the
the lower surface; (b) generation and Gaussian fit of the aperture distribution.
case for LBM, where the inward distribution functions to the in-
tegration domain need to be determined before streaming on the
Bhatnagar-Gross-Krook (LBGK) approximation (Qian et al., 1992): boundaries. Therefore, appropriate equations are needed for calcu-
lating those distribution functions on the boundaries for a given
1
fi (x + ei t, t + t ) − fi (x, t ) = − [ fi (x, t ) − fieq (x, t )] (6) boundary condition. A number of boundary treatment methods
τ have been proposed, such as the bounce-back scheme (He et al.,
where the left hand side of the equation represents the stream- 1997), the extrapolation scheme (Guo et al., 2002), the utilization
ing step and the right hand side represents the collision in the of the immersed boundary treatments (Fadlun et al., 20 0 0), and
evolution of the distribution function, fi eq (x, t) is the equilibrium the reconstruction of the distribution functions from the density,
distribution function, and τ is the relaxation parameter that gov- velocity, and the rate of strain (Latt et al., 2008). In this study,

D3Q19 model in LBM

15
Inl A
et
pre
No-slip boundary ssu
5 re
12
11 B

16 3
8 7

0
2 1
No-slip boundary
10 B'
9
4 18
Ou
tle ion
13 tp
14 re s ect
6 A' sur dir
e w
Flo
Geometry model of the fracture
17

Fig. 7. Illustration of the numerical model for flow simulation through a 3D self-affine fracture.
380 M. Wang et al. / Advances in Water Resources 96 (2016) 373–388

Table 2 4. Simulation of nonlinear fluid flow through 3D self-affine

Dimensional mapping between the physical units and LBM simulation
fractures
units.

Unit Symbol Dimensions Mapping 4.1. Model setup

Length L [L] Lm =xph /xLB
Density ρ [M][L]−3 ρ m =ρ ph /ρ LB For realistic modelling of the flow through rough fractures, the
Kinematic viscosity υ [L]2 [T]−1 υ m =υ ph /υ LB numerically generated fractures should be representative of the ge-
Time t [T] Tm =Lm 2 /υ m ometries of natural fractures, such as the surface fractal dimen-
Mass m [M] Mm =ρ m Lm 3
sions, surface asperity height distributions and aperture distribu-
Velocity u [L][T]−1 Um = Lm /Tm
Force F [M][L][T]−2 Fm =M m Lm /Tm 2 tions (Ogilvie et al., 2006). As stated in Section 2.1, natural rock
Pressure P [M][L]−1 [T]−2 Pm = Mm /(Lm Tm 2 ) fractures generally possess a Hurst exponent H ranging between
0.45 and 0.85, and therefore, four fractures with H = 0.5, 0.6, 0.7
and 0.8 were created for numerical simulations, with the morphol-
the non-equilibrium extrapolation method developed by Guo et al. ogy of the lower surfaces shown in Fig. 1. To properly model the
(2002) for curved no-slip boundary conditions was used to cal- secondary roughness of the fracture surfaces, a resolution of 25 μm
culate the missing components of the distribution function at the in x, y and z directions was taken for the lattice model, which
nodes neighboring to the solid nodes. The extrapolation method is can realistically be obtained by micro-CT scanners (Crandall et al.,
of second-order accuracy and has a better numerical stability com- 2010). This resolution generally guarantees a good balance between
pared with the bounce-back based treatments for curved bound- the accuracy for roughness characterization and the computational
aries; it is simple and can be extended to include velocity, temper- cost in the numerical simulations. To save computational cost, the
ature, and pressure boundary conditions and their derivatives (Guo fracture size was chosen as 50 mm long and 25 mm wide. The
et al., 2002). mean and the standard deviation of the aperture were taken as
How to relate the non-dimensional LB world to the real dimen- b0 = 0.5 mm and σ = 0.15 mm, with a moderate value of σ /b0 = 0.3.
sions of hydrogeological flow problems is another issue that merits Such an aperture configuration may reduce the effect of contact
attention. In the LB simulations, both the lattice spacing and the area (zero-aperture region), given that the main purpose of this
time step are usually non-dimensional and typically set to unity. study is to examine the effect of surface roughness on the nonlin-
For a simulation to be of practical use, the following mappings be- ear flow through the fractures.
tween the LBM units and the physical units are required: To evaluate the influence of the primary waviness and sec-
ondary roughness on the flow behaviors in the fractures, two
x ph t ph m ph kinds of models were established for each fracture: one contains
Lm = , Tm = , Mm = (12)
xLB tLB mLB the secondary roughness (denoted by Frc1−Frc4) and the other
where Lm , Tm and Mm are the length, time and mass mapping, re- only considers the decomposed primary waviness (denoted by
spectively, and xγ , tγ and mγ (γ = ph or LB) represent the length, Frc1#−Frc4#). A total of 14 pressure gradients (-∇ P) varying from
time and mass in the physical or LBM systems. 1 to 50 0 0 Pa/m were applied between the inlet and outlet of each
The length mapping Lm is determined by the choice of res- fracture model, resulting in a total of 112 flow simulations. The
olution of the simulation. For instance, if the lattice resolution fracture walls and the lateral sides of the fractures were modeled
is chosen as 100 nodes for 1-mm mean aperture for the flow as no-slip velocity boundaries and the fractures were assumed to
modeling, the physical spacing (xph ) for one lattice will be be non-deformable. Fig. 7 illustrates the boundary conditions and
1×10−5 m, i.e. xph = 1×10−5 m. The LB simulation usually takes a the lattice mesh for the simulation of flow in the fracture model
constant spacing of unity as default, i.e. xLB = 1, and hence one Frc2. The numerical simulations were performed for water with a
has Lm = xph /xLB = 1×10−5 m. density of 998.2 kg/m3 and a dynamic viscosity of 10−3 Pa s, with
The mass mapping Mm can be obtained from the density of the gravitational effect being neglected. At each pressure gradient,
fluid. The LB simulation is usually initialized with ρ LB = 1 at all the velocity and pressure in the fractures were initialized to zero.
nodes. Noticing the dimensions of density [M][L]−3 and ρ py = Mm The flow was considered to reach a steady state and the corre-
Lm −3 ρ LB yields sponding flow rate was recorded for further analysis, when the tol-
erance defined as |ui +1 −ui | ≤ 1 × 10−5 m/s is satisfied, where u is
ρ ph 3 the flow velocity and i is the time step.
Mm = L (13)
ρLB m To check for the mesh (lattice) independence, two numerical
The time mapping Tm is imposed by the choice of the fluid simulations with lattice resolutions of 25 and 12.5 μm, respec-
kinematic viscosity. In the LBGK model, the viscosity is related to tively, were performed on the Frc2 model at the pressure gradient
the relaxation parameter τ and the lattice pseudo-sound-speed cs , -∇ P = 50 0 0 Pa/m. The steady-state flow rate of the two simulations
υ LB = (τ −0.5)cs 2 tLB , with dimensions of [L]2 [T]−1 , which yields is 7.65×10−7 m3 /s and 7.67×10−7 m3 /s, respectively, with a small
relative difference of 0.26%. This implies that a lattice resolution of
υLB 2
Tm = L (14) 25 μm is sufficiently accurate for the flow simulations, and a finer
υ ph m resolution only has a limited effect on accuracy, but drastically in-
The mappings of length, time and mass provide the base to creases the computational cost.
convert other dimensional mappings. A summary of the dimen-
sional mappings between the LB simulation units and the physical 4.2. Numerical results of fluid flow in the fractures
units is listed in Table 2. Note that the definition of the lattice vis-
cosity implies that the relaxation parameter should be selected as Without loss of generality, the numerical results in the fracture
τ > 0.5. models Frc2 and Frc2# were taken for illustration and comparison.
Based on the above theory, a LBM code using the D3Q19 lattice Figs. 8 and 9 show the 3D distributions of pressure and velocity
model, named as 3DFracFlow, was developed to recover the solu- in Frc2 and Frc2# at the pressure gradient of 50 0 0 Pa/m. One ob-
tion of the NS equations for the incompressible fluid flow in rock serves from Fig. 8 that the pressure varies almost linearly along
fractures, and it has been verified for water flow through a cylin- the flow direction (x-direction) in both Frc2 and Frc2# models. The
drical tube (see Appendix). variation of pressure along the direction perpendicular to flow (y-
 M. Wang et al. / Advances in Water Resources 96 (2016) 373–388 381

Fig. 8. Three-dimensional pressure field (unit: Pa) at −∇ P = 50 0 0 Pa/m in (a) model Frc2 and (b) model Frc2#.

Fig. 9. Three-dimensional fracture velocity field (unit: m/s) at −∇ P = 50 0 0 Pa/m in (a) model Frc2 and (b) model Frc2#.

direction) is caused by the local changes of aperture. The similar interval y = [0.020 m, 0.022 m] along section B-B’ (x = 0.024 m, see
distributions of pressure in Frc2 and Frc2# indicate that the local Fig. 7) is plotted, the backflow regions in Frc2 at the pressure gra-
asperities of the secondary roughness have a limited influence on dient of 50 0 0 Pa/m constitute the other contribution of the small-
the overall pressure distribution in the fractures. scale unevenness in reducing the effective advection area, which
Fig. 9 shows that the flow is channeled into main high-speed obviously could not be revealed in 2D fracture flow simulations.
streams with many “dead” zones of very low velocities, with a The tortuosity of flow paths in 3D fractures may effectively
maximum velocity of about 0.29 m/s in Frc2# and 0.38 m/s in Frc2, reflect the complexity of the velocity field induced by the small
respectively. It can also be inferred from Fig. 9 that the velocity and/or large scale roughness. Fig. 12 compares the particle stream-
field in Frc2 is more complex than that in Frc2#, with the regions lines in Frc2 and Frc2# at the pressure gradient of 50 0 0 Pa/m, in-
of high velocity being more discretely distributed or disordered, dicating that the streamlines in Frc2 are much more tortuous than
which is similar to the numerical results in 2D fracture models those in Frc2#. In this study, the definition given by Brown et al.
(Zou et al., 2015). The increase of high velocity regions in Frc2 is (1998) is adopted for quantitatively evaluating the tortuosity:
mostly contributed by the local effects of the abruptly changed as- 2
La
perities of the secondary roughness, while the low velocity regions τ 2= (15)
in both Frc2 and Frc2# are mostly controlled by the undulations of L
waviness, where the eddy flow may occur as the pressure gradient where La is the actual length that the fluid traverses and L is de-
increases. fined as the macroscopic length of the fracture models (L = 50 mm
To further illustrate the influence of the secondary roughness in this study). Based on the simulation results, La could be practi-
on the generation of eddy flow, Fig. 10 compares the velocity cally determined by exporting the position information of the par-
contours (x- and z-components) and streamlines at the interval ticles from Tecplot 360 and summing the travel distances. The par-
x = [0.013 m, 0.016 m] along section A-A’ (y = 0.0125 m, see Fig. 7) in ticles that did not reach the outlet (e.g. particles deposited on the
the Frc2 and Frc2# models under different pressure gradients. The fracture walls) were excluded, and a number of over 400 effective
results indicate that with the increase of pressure gradient from particles were tracked for obtaining the mean value of La in each
50 Pa/m (Fig. 10a) to 50 0 0 Pa/m (Fig. 10c), the flow in Frc2 devi- model.
ates much from linearity, with the eddy flow regions being gener- Fig. 13 shows the variations of the tortuosity τ 2 with the pres-
ated in the vicinities of the locally changed asperities. In the Frc2# sure gradient in the Frc2 and Frc2# models, respectively, show-
model where the secondary roughness is not accounted, however, ing a higher value of tortuosity at various pressure gradients in
the pattern of streamlines at high pressure gradient (Fig. 10d) is Frc2 where the secondary roughness makes more tortuous the flow
almost the same with that at low pressure gradient (Fig. 10b). paths in the vicinities of the sharp-cornered asperities. The results
As the pressure gradient increases, the flow in Frc2 becomes demonstrate that it is the secondary roughness (small-scale un-
more turbulent, and the effective advection area in the y-direction evenness) that increases the tortuosity of flow paths and enhances
becomes narrower due to the expansion of the eddy flow regions. the local complexity of velocity distributions with eddy and back
As evidenced in Fig. 11 where the x-component velocity at the flow regions. The primary roughness (large-scale waviness), on the
382 M. Wang et al. / Advances in Water Resources 96 (2016) 373–388

Fig. 10. Streamlines and velocity distribution in the interval x = [0.013 m, 0.016 m] at section A-A’ (y = 0.0125 m) under different pressure gradient. (a) Model Frc2, -∇ P
=50 Pa/m, (b) model Frc2#, -∇ P =50 Pa/m, (c) model Frc2, -∇ P =50 0 0 Pa/m, and (d) model Frc2#, -∇ P =50 0 0 Pa/m.

Fig. 11. x-component velocity in the interval y = [0.02 m, 0.022 m] at section B-B’ (x = 0.024 m) at -∇ P = 50 0 0 Pa/m. (a) Model Frc2 and (b) model Frc2#.

Fig. 12. Particle pathlines at -∇ P =50 0 0 Pa/m. (a) Model Frc2 with secondary roughness and (b) model Frc2# without secondary roughness. The color on the streamlines
represents the magnitude of velocity. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

other hand, largely determines the pressure distribution and initial and the localized eddy formation due to the inertial effects of flow
tortuosity. (Chen et al., 2015; Zhou et al., 2015). The nonlinearity of flow in
fractures can be induced by the increase of pressure gradient or by
the surface roughness. The most extensively-adopted mathematical
4.3. Influence of surface roughness on the nonlinear flow behavior
description of the nonlinear flow in rough-walled fractures is the
Forchheimer’s law (Bear, 1972), written as
It has been understood that the flow in fractures may deviate
from linearity as a result of the changes in flow velocity or direc-
tion along the flow paths due to contact regions or obstructions −∇ P = aQ + bQ 2 (16)
 M. Wang et al. / Advances in Water Resources 96 (2016) 373–388 383

Table 3
Summary of the best-fitted coefficients a and b in the Forch-
heimer equation, critical Reynolds numbers, and critical hy-
draulic gradient.

Model no. a (kg·s−1 ·m−5 ) b (kg·m−8 ) Rec Jc

Frc1 1.49 × 1010

1.75 × 1016
1.8 0.07
Frc2 5.45 × 109 1.48 × 1015 7.7 0.11
Frc3 4.64 × 109 6.55 × 1014 14.9 0.19
Frc4 4.30 × 109 4.10 × 1014 22.0 0.26
Frc1# 4.46 × 109 2.45 × 1014 38.2 0.46
Frc2# 4.25 × 109 2.17 × 1014 41.1 0.47
Frc3# 4.14 × 109 1.97 × 1014 44.1 0.49
Frc4# 4.08 × 109 1.87 × 1014 45.7 0.50

a and b being listed in Table 3. Both the values of a and b de-

crease with the decrease of surface roughness, but vary in much
larger ranges in the models Frc1−Frc4, indicating that the sec-
ondary roughness plays an important role in both the intrinsic per-
Fig. 13. Variation of tortuosity with pressure gradient in the models Frc2 and meability (represented by a) and the nonlinear flow behavior (rep-
Frc2#. resented by b), especially for rougher fractures with H < 0.6−0.7.
Fig. 15 plots the nonlinear effect factor, α , calculated by
Eq. (19) against the pressure gradient (−∇ P) or hydraulic gradient
where Q is the volumetric flow rate, −∇ P is the pressure gradi-
(J), showing that α increases much faster with J and deviates much
ent, a= 12μ3 and b= βμ
2 2
are the linear and nonlinear coefficients,
weh w e
h from each other in the models Frc1−Frc4, while it evolves almost
respectively, β is called the Forchheimer coefficient, w is the frac- in the same pattern in Frc1#−Frc4#. If, as commonly done in the
ture width perpendicular to the flow direction, eh is the hydraulic literature (Zhou et al., 2015), a critical value of α = 5% is accepted
aperture, and μ is the fluid viscosity. to quantify the transition of flow regimes from linear to nonlinear,
When the inertial effect is negligible at low velocity (i.e. β = 0), then the critical Reynolds number (Rec ) and the critical hydraulic
the nonlinear term (bQ2 ) vanishes and the Forchheimer’s law re- gradient (Jc ) could be obtained, as shown in Fig. 15 and Table 3.
duces to the linear Darcy’s law. For fluid flow in fractures, the The results show much smaller magnitudes of Rec (1.8−22.0) and
Reynolds number (Re), defined as the ratio of inertial forces to vis- Jc (0.07−0.26) in the models Frc1−Frc4 than those in the ranges of
cous forces, is commonly adopted to determine the onset of non- 38.2−45.7 and 0.46–0.50, respectively, in Frc1#−Frc4#, indicating
linearity: that the secondary roughness significantly enhances the nonlinear
ρveh ρ Q flow through fractures and leads to earlier onset of nonlinearity.
Re = = (17)
μ μw
4.4. Influence of surface roughness on transmissivity, hydraulic
When the flow follows the Forchheimer’s law, a critical aperture and tortuosity
Reynolds number can be obtained from Eq. (17) (Javadi et al., 2014;
Zhou et al., 2015): Transmissivity (T) is an important hydraulic property for charac-
aρα terizing the macroscopically observed flow resistances in fractures.
Rec = (18)
bμw(1 − α ) When Re is low (less than 1), the intrinsic transmissivity (T0 ) is
commonly regarded as a constant independent of Re and can be
where α is the nonlinear effect factor defined as the ratio of the determined using the best-fitted values of the linear coefficient a
nonlinear pressure drop to the overall pressure drop (Zimmerman (i.e. T0 = μ/a). As Re increases, the apparent transmissivity (Ta ) can
et al., 2004; Zeng and Grigg., 2006): be used to characterize the nonlinear flow by extending the lin-
bQ 2 ear Darcy’s law with Ta = μQ/(-∇ P). The normalized transmissivity
α= (19) (Ta /T0 ) can then be determined by Eqs. (16) and (17) (Zimmerman
aQ + bQ 2
et al., 2004):
Note that the Reynolds number defined in Eq. (17) is constant
Ta 1
at every cross-section of the fractures, or in other words, the crit- = (20)
ical flow rate (Qc ) through the fractures and the critical pressure T0 1 + cRe
gradient (−∇ Pc ) are independent of the choice of the length scale, where c is a dimensionless coefficient determined by the coeffi-
provided that the critical Reynolds number is known. A critical hy- cients a and b in the Forchheimer’s equation as c = bwμ/aρ .
draulic gradient (Jc ) defined as Jc = −∇ Pc /ρ g was therefore intro- Table 4 lists the values of coefficient c and intrinsic transmis-
duced by Liu et al. (2016a) as a more practical parameter for de- sivity T0 for each model, showing that the value of c in the mod-
tecting the onset of nonlinear flow in fractured media. els Frc1−Frc4 decreases by about one order of magnitude with
Fig. 14 plots the relationship between the pressure gradient the increase of the Hurst exponent from 0.50 to 0.80, while it re-
(−∇ P) and volumetric flow rate (Q) for fluid flow through the frac- duces in a rather narrow range in Frc1#−Frc4#. Fig. 16 plots the
ture models (Frc1−Frc4 and Frc1#−Frc4#) with H = 0.5, 0.6, 0.7 and normalized transmissivity against the Reynolds number, indicat-
0.8, respectively. The curves show that the nonlinear flow does oc- ing that the numerical data fit fairly well with the predictions by
cur in the fractures no matter whether the secondary roughness Eq. (20). It can also be observed that the normalized transmissiv-
is accounted or not, but the degree of nonlinearity increases when ity remains a constant of 1 for Re < 1 except for the model Frc1
the Hurst exponent becomes smaller or the secondary roughness with the highest surface roughness. As Re increases, the normal-
turns to play a role. It was found that the numerically obtained ized transmissivity decreases dramatically, which further confirms
curves between −∇ P and Q can be well fitted with the Forch- the deviation of flow from linearity. At the same Reynolds num-
heimer’s equation, with the best-fitted values of the coefficients ber (Re > 1), the nonlinear deviation in Frc1−Frc4 is more signifi-
384 M. Wang et al. / Advances in Water Resources 96 (2016) 373–388

Fig. 14. Numerically-obtained and best-fitted pressure versus flow rate curves for the models Frc1−Frc4 and Frc1#−Frc4#.

Table 4
Summary of the best-fitted coefficients c in Eq. (20), intrinsic
transmissivity, intrinsic hydraulic aperture and mean tortuosity.

Model no. c T0 (mm4 ) eh0 (mm) τ2

Frc1 2.95 × 10−2 6.73 × 10−5 0.318 1.162

Frc2 6.80 × 10−3 1.84 × 10−4 0.445 1.096
Frc3 3.54 × 10−3 2.16 × 10−4 0.470 1.075
Frc4 2.39 × 10−3 2.33 × 10−4 0.482 1.054
Frc1# 1.38 × 10−3 2.25 × 10−4 0.476 1.060
Frc2# 1.28 × 10−3 2.38 × 10−4 0.485 1.050
Frc3# 1.19 × 10−3 2.42 × 10−4 0.488 1.037
Frc4# 1.15 × 10−3 2.51 × 10−4 0.494 1.027

Fig. 17 shows the variation of the ratio (eh /em ) of the equivalent
hydraulic aperture to the mean mechanical aperture (em = 0.5 mm
in this study) with the pressure gradient and hydraulic gradient.
The results show that for J < 0.01 or -∇ P < 100 Pa/m, the ratio
eh /em almost remains constant, indicating that the flow is lami-
nar and the cubic law applies. An intrinsic hydraulic aperture (eh0 )
Fig. 15. Evolution of non-linear effect factor α with pressure gradient and hydraulic
could hence be calculated with the cubic law in this flow regime,
gradient. as listed in Table 4. The intrinsic hydraulic apertures in the mod-
els Frc1−Frc4 varies in a wide range from 64 to 96% of their
mechanical apertures, but in a narrow range from 95 to 99% in
cant than that in Frc1#−Frc4# due to the effect of the secondary Frc1#−Frc4#. This difference of hydraulic aperture results purely
roughness. from the secondary roughness, and it deceases significantly as the
The variation of transmissivity is essentially related to the roughness exponent H increases. As J increases in the range be-
change of the equivalent hydraulic aperture (eh ), which could be tween 0.01 and 0.1, eh /em decreases gently and the flow enters a
estimated with the cubic law using the total flow rate through the weak inertial regime, as a combined effect of the inertial term of
fracture and the pressure gradient between the inlet and outlet. the NS equations and the fracture surface geometry. With a further
 M. Wang et al. / Advances in Water Resources 96 (2016) 373–388 385

Fig. 16. Variation of normalized transmissivity (Ta /T0 ) with Reynolds number (Re).

Fig. 18. Tortuosity (τ 2 ) as a function of the Hurst exponent. The number in the
parentheses denotes the number of streamlines. (a) Models Frc1−Frc4 and (b) mod-
els Frc1#−Frc4#.
Fig. 17. Variation of the ratio of equivalent hydraulic aperture to mean mechanical
aperture with pressure gradient and hydraulic gradient.

increase of J over 0.1, a more notable decrease of eh /em , especially low-frequency component (large-scale waviness or primary rough-
in the models Frc1−Frc4, shows that the flow may have transited ness) and a high-frequency one (small-scale unevenness or sec-
into a strong inertial regime, where the effective advective flow ondary roughness) using wavelet analysis. The lattice Boltzmann
area further decreases with the increasing boundary layers formed method was adopted and a 3D LBM code was developed for mod-
by eddy flow and backflow regions. elling the fluid flow in the fractures. It was demonstrated by the
Fig. 18 shows a boxplot of the tortuosity τ 2 calculated by numerical results that the primary roughness controls the pres-
Eq. (15) as a function of the roughness exponent H. The tortuos- sure distribution and the main flow directions, with low velocity
ity deceases almost linearly in the models Frc1#−Frc4#, but not regions occurring in the boundary layers. The sharply changed as-
in Frc1−Frc4. Table 4 lists the mean values of tortuosity, showing perities of the secondary roughness enhance the local complexity
that the tortuosity vary with the degree of roughness in a higher of the velocity distributions by generating and expanding the eddy
range between 1.054 and 1.162 in Frc1−Frc4 and in a lower range flow and back flow regions, and render the flow paths more tortu-
between 1.027 and 1.060 in Frc1#−Frc4#. The tortuosity of flow ous. The Forchheimer’s law proves to well represent the nonlinear
paths becomes more manifested in rougher fractures with lower flow behaviors in the 3D fractures. The inertial effect induced by
value of H, where the sharp-shaped asperities of surface roughness the primary roughness differs only marginally in the fractures of
render the flow in the local vicinities more turbulent. varying roughness exponent (H), while the secondary roughness
plays a dominating role in the occurrence of nonlinearity, espe-
5. Conclusions cially for rougher fractures with H < 0.6−0.7. The apparent trans-
missivity and the equivalent hydraulic aperture of the fractures
This primary goal of this study is to examine the influence of are influenced not only by the pressure gradient (or the Reynolds
the multi-scale surface roughness on the fluid flow behaviors in 3D number), but also significantly by the secondary roughness. The
self-affine rock fractures through high-resolution numerical sim- tortuosity of flow paths, though varies gently with the pressure
ulations. The fracture surface roughness was decomposed into a gradient, increases notably with the surface roughness.
386 M. Wang et al. / Advances in Water Resources 96 (2016) 373–388

To the best of our knowledge, this study presents for the first Appendix. Verification of the LBM code
time such a detailed and quantitative analysis on the nonlinear
flow behaviors in 3D rough fractures induced by multi-scale sur- The Hagen-Poiseuille flow in a cylindrical tube was simulated
face roughness. The 3D numerical simulations performed in this to verify the LBM code, 3DFracFlow, developed on the basis of the
study have advantages over previous 2D simulations (Madadi et al., D3Q19 lattice model. The x-directional velocity (ux ) of the flow is
2003; Zou et al., 2015) for elaborating the mechanisms of nonlin- known as:
ear flow caused by the secondary roughness, such as the back flow
phenomenon and the tortuosity of flow paths around the low aper-  
ture regions. Given the main purpose of this study, no attempt is 1 D2 r2
ux = −∇ Px − (A.1)
made to evaluate the effects of the void space geometry (e.g. the μ 16 4
aperture distribution with contact regions) on the nonlinear flow
in 3D fractures. Our future study will focus on the combined ef-
fects of the surface roughness and aperture variation, the scale ef- where −∇ Px is the pressure gradient, D is the diameter of the
fects and the anisotropic behaviors of rough fractures as well as cylindrical tube, and r is the radial distance from the centerline
the nonlinear flow behaviors in rough fracture networks. of the tube.
Without loss of generality, the flow in a circular tube of 10 mm
in diameter and 40 mm in length was simulated under the pres-
sure gradient of 10−6 g/(mm2 ·s2 ). The relaxation time and the den-
sity of the fluid were taken as 0.8 and 1 in lattice units, respec-
tively. The lattice was created at the resolution of 20 points along
Acknowledgments the radial direction (10 mm), resulting in a total of 35,721 points
in the model. Fig. A.1 shows the distribution of the calculated x-
The authors gratefully thank Graham C. Sander at Loughbor- directional velocity (ux ) in the tube, indicating that the numerical
ough University and the anonymous reviewers for their kind ef- result agrees perfectly with the analytical solution. Fig. A.2 plots,
forts and constructive comments in improving this work. The fi- respectively, the relative error of ux between the analytical and nu-
nancial supports from the National Natural Science Foundation of merical solutions along the radial direction and the convergence
China (Nos. 51579188 and 51409198) are gratefully acknowledged. rate of the LBM code, demonstrating a high accuracy of the numer-
The first author would like to thank the financial support from the ical solution (with the relative error lower than 0.3%) and a good
program of China Scholarship Council (CSC). numerical stability of the algorithm.

Fig. A.1. Distribution of velocity ux in Poiseuille flow: (a) the contour of the velocity ux at the slice of y=0, (b) at the slice of x=0.5 L, and (c) on the line of x=0.5 L, y=0.
 M. Wang et al. / Advances in Water Resources 96 (2016) 373–388 387

Fig. A.2. (a) Relative error of ux between the analytical and numerical solutions along the radial direction; (b) residual error versus time step.

References Eker, E, Akin, S, 2006. Lattice Boltzmann simulation of fluid flow in synthetic frac-
tures. Transport Porous Media 65 (3), 363–384.
Anwar, S, Sukop, MC, 2009. Lattice Boltzmann models for flow and transport in sat- Fadlun, EA, Verzicco, R, Orlandi, P, et al., 20 0 0. Combined immersed-boundary
urated karst. Groundwater 47 (3), 401–413. finite-difference methods for three-dimensional complex flow simulations. J.
Auradou, H., 2009. Influence of wall roughness on the geometrical, mechanical and Comput. Phys. 161 (1), 35–60.
transport properties of single fractures. J. Phys. D 42 (21), 214015. Filippova, O, Hänel, D, 1998. Boundary-fitting and local grid refinement for lat-
Babadagli, T, Ren, X, Develi, K, 2015. Effects of fractal surface roughness and lithol- tice-BGK models. Int. J. Modern Phys. C 9 (08), 1271–1279.
ogy on single and multiphase flow in a single fracture: an experimental inves- Gao, Y, Wong, LNY, 2015. A modified correlation between roughness parameter Z2
tigation. Int. J. Multiph. Flow 68, 40–58. and the JRC. Rock Mech. Rock Eng. 48 (1), 387–396.
Barton, N, Choubey, V, 1977. The shear strength of rock joints in theory and practice. Ginzburg, I., 2013. Multiple anisotropic collisions for advection–diffusion lattice
Rock Mechanics 10 (1-2), 1–54. Boltzmann schemes. Adv. Water Resour. 51, 381–404.
Bear, J., 1972. Dynamics of Fluids in Porous Media. Guo, Z, Zheng, C, Shi, B, 2002. An extrapolation method for boundary conditions in
Belem, T, Homand-Etienne, F, Souley, M, 20 0 0. Quantitative parameters for rock lattice Boltzmann method. Phys. Fluids 14 (6), 2007–2010.
joint surface roughness. Rock Mech. Rock Eng. 33 (4), 217–242. Gutfraind, R., Hansen, A., 1995. Study of fracture permeability using lattice gas au-
Benzi, R, Succi, S, Vergassola, M, 1992. The lattice Boltzmann equation: theory and tomata. Transport Porous Media 18 (2), 131–149.
applications. Phys. Rep. 222 (3), 145–197. He, X, Luo, LS, 1997. Lattice Boltzmann model for the incompressible Navier–Stokes
Berkowitz, B., 2002. Characterizing flow and transport in fractured geological media: equation. J. Statis. Phys. 88 (3–4), 927–944.
a review. Adv. Water Resour. 25 (8), 861–884. He, X, Zou, Q, Luo, LS, et al., 1997. Analytic solutions of simple flows and analysis
Boffa, JM, Allain, C, Hulin, JP, 1998. Experimental analysis of fracture rugosity in of nonslip boundary conditions for the lattice Boltzmann BGK model. J. Statis.
granular and compact rocks. Eur. Phys. J. Appl. Phys. 2 (3), 281–290. Phys. 87 (1–2), 115–136.
Briggs, S, Karney, BW, Sleep, BE, 2014. Numerical modelling of flow and transport Indraratna, B, Kumara, C, Zhu, SP, Sloan, S, 2014. Mathematical modeling and ex-
in rough fractures. J. Rock Mech. Geotech. Eng. 6 (6), 535–545. perimental verification of fluid flow through deformable rough rock joints. Int.
Brown, S, Caprihan, A, Hardy, R, 1998. Experimental observation of fluid flow chan- J. Geomech. 15 (4), 04014065.
nels in a single fracture. J. Geophys. Res. 103 (B3), 5125–5132. ISRM, 1978. Suggested methods for the quantitative description of discontinuities in
Brown, SR, Scholz, CH, 1985. Broad bandwidth study of the topography of natural rock masses. Int. J. Rock Mech. Mining Sci. 15, 319–368.
rock surfaces. Geophys. Res. Lett. 90 (B14), 12575–12582. Javadi, M, Sharifzadeh, M, Shahriar, K, Mitani, Y, 2014. Critical Reynolds number
Brush, DJ, Thomson, NR, 2003. Fluid flow in synthetic rough-walled fractures: for nonlinear flow through rough-walled fractures: the role of shear processes.
Navier–Stokes, Stokes, and local cubic law simulations. Water Resour. Res. 39 Water Resour. Res. 50 (2), 1789–1804.
(4), 1085. Jeong, N, Choi, DH, Lin, CL, 2006. Prediction of Darcy–Forchheimer drag for micro–
Chae, BG., Ichikawa, Y, Jeong, GC, Seo, YS., Kim, BC., 2004. Roughness measure- porous structures of complex geometry using the lattice Boltzmann method. J.
ment of rock discontinuities using a confocal laser scanning microscope and Micromech. Microeng. 16 (10), 2240.
the Fourier spectral analysis. Eng. Geol. 72 (3), 181–199. Jing, L, Nordlund, E, Stephansson, O, 1992. An experimental study on the anisotropy
Chamoli, A, Bansal, AR, Dimri, VP, 2007. Wavelet and rescaled range approach for and stress-dependency of the strength and deformability of rock joints. Int. J.
the Hurst coefficient for short and long time series. Comput. Geosci. 33 (1), Rock Mech. Mining Sci. Geomech. Abstracts 29 (6), 535–542.
83–93. Kang, Q, Zhang, D, Chen, S, 2002. Unified lattice Boltzmann method for flow in mul-
Chen, L, Kang, Q, Mu, Y, et al., 2014. A critical review of the pseudopotential mul- tiscale porous media. Phys. Rev. E 66 (5), 056307.
tiphase lattice Boltzmann model: methods and applications. Int. J. Heat Mass Khoshelham, K., Altundag, D, Ngan-Tillard, D, Menenti, M., 2011. Influence of range
Transfer 76, 210–236. measurement noise on roughness characterization of rock surfaces using terres-
Chen, S, Doolen, GD, 1998. Lattice Boltzmann method for fluid flows. Annu. Rev. trial laser scanning. Int. J. Rock Mech. Mining Sci. 48 (8), 1215–1223.
Fluid Mech. 30 (1), 329–364. Kim, I, Lindquist, WB, Durham, WB, 2003. Fracture flow simulation using a finite-d-
Chen, S, Martinez, D, Mei, R, 1996. On boundary conditions in lattice Boltzmann ifference lattice Boltzmann method. Phys. Rev. E 67 (4), 046708.
methods. Phys. Fluids 8 (9), 2527–2536. Latt, J, Chopard, B, Malaspinas, O, Deville, M, Michler, A, 2008. Straight velocity
Chen, YF, Zhou, JQ, Hu, SH, Hu, R, Zhou, CB, 2015. Evaluation of Forchheimer equa- boundaries in the lattice Boltzmann method. Phys. Rev. E 77 (5), 056703.
tion coefficients for non-Darcy flow in deformable rough-walled fractures. J. Hy- Lee, SH., Lee, KK., Yeo, IW., 2014. Assessment of the validity of Stokes and Reynolds
drol. 529, 993–1006. equations for fluid flow through a rough-walled fracture with flow imaging.
Crandall, D, Bromhal, G, Karpyn, ZT, 2010. Numerical simulations examining the re- Geophys. Res. Lett. 41 (13), 4578–4585.
lationship between wall-roughness and fluid flow in rock fractures. Int. J. Rock Li, Y, Chen, YF, Zhou, CB, 2014. Hydraulic properties of partially saturated rock frac-
Mech. Mining Sci. 47 (5), 784–796. tures subjected to mechanical loading. Eng. Geol. 179, 24–31.
Dardis, O, McCloskey, J, 1998. Lattice Boltzmann scheme with real numbered solid Liu, HH., Bodvarsson, GS., Lu, S, Molz, FJ., 2004. A corrected and generalized suc-
density for the simulation of flow in porous media. Physical Review E 57 (4), cessive random additions algorithm for simulating fractional Levy motions.
4834. Mathem. Geol. 36 (3), 361–378.
De Rosis, A., 2014a. A lattice Boltzmann model for multiphase flows interacting with Liu, RC, Li, B, Jiang, YJ, 2016a. Critical hydraulic gradient for nonlinear flow through
deformable bodies. Adv. Water Resour. 73, 55–64. rock fracture networks: the roles of aperture, surface roughness, and number of
De Rosis, A., 2014b. A lattice Boltzmann-finite element model for two-dimensional intersections. Adv. Water Resour. 88, 53–65.
fluid–structure interaction problems involving shallow waters. Adv. Water Re- Liu, RC, Li, B, Jiang, YJ, Huang, N, 2016b. Review: Mathematical expressions for esti-
sour. 65, 18–24. mating equivalent permeability of rock fracture networks. Hydrogeol. J. 1–27.
Develi, K., Babadagli, T., 1998. Quantification of natural fracture surfaces using frac- Llewellin, EW, 2010. LBflow: an extensible lattice Boltzmann framework for the sim-
tal geometry. Mathem. Geol. 30 (8), 971–998. ulation of geophysical flows. Part I: theory and implementation. Comput. Geosci.
Dou, Z, Zhou, Z, Sleep, BE, 2013. Influence of wettability on interfacial area dur- 36 (2), 115–122.
ing immiscible liquid invasion into a 3D self-affine rough fracture: lattice Boltz- Madadi, M, Sahimi, M, 2003. Lattice Boltzmann simulation of fluid flow in fracture
mann simulations. Adv. Water Resour. 61, 1–11. networks with rough, self-affine surfaces. Phys. Rev. E 67 (2), 026309.
388 M. Wang et al. / Advances in Water Resources 96 (2016) 373–388

Madadi, M, VanSiclen, CD, Sahimi, M, 2003. Fluid flow and conduction in two-di- Shapiro, S, Wilk, M, 1965. An analysis of variance test for normality (complete sam-
mensional fractures with rough, self-affine surfaces: a comparative study. J. ples). Biometrika 52 (3), 591–611.
Geophys. Res. 108 (B8). Simonsen, I, Hansen, A, Nes, OM, 1998. Determination of the Hurst exponent by use
Mallat, SG., 1989. A theory for multiresolution signal decomposition: the wavelet of wavelet transforms. Phys. Rev. E 58 (3), 2779.
representation. Pattern Anal. Mach. Intell. IEEE Trans. 11 (7), 674–693. Succi, S, Benzi, R, Higuera, F, 1991. The lattice Boltzmann equation: a new tool for
Mandelbrot Benoit, B., 1983. The fractal geometry of nature. W.H. Freeman, New computational fluid-dynamics. Phys. D 47 (1–2), 219–230.
York. Succi, S., 2001. The Lattice Boltzmann Equation: For Fluid Dynamics and Beyond.
Mehrishal, A, Sharifzadeh, M, 2013. Evaluation of the hydraulic aperture of a rock Oxford university press.
joint using wavelet theory. Geosyst. Eng. 16 (1), 119–127. Sukop, MC., Huang, H, Alvarez, PF, Variano, EA, Cunningham, KJ., 2013. Evaluation
Mei, R, Shyy, W, Yu, D, Luo, LS, 20 0 0. Lattice Boltzmann method for 3-D flows with of permeability and non-Darcy flow in vuggy macroporous limestone aquifer
curved boundary. J. Comput. Phys. 161 (2), 680–699. samples with lattice Boltzmann methods. Water Resour. Res. 49 (1), 216–230.
MŁynarczuk, M., 2010. Description and classification of rock surfaces by means of Walsh, R, McDermott, C, Kolditz, O, 2008. Numerical modeling of stress-permeabil-
laser profilometry and mathematical morphology. Int. J. Rock Mech. Mining Sci. ity coupling in rough fractures. Hydrogeol. J. 16 (4), 613–627.
47 (1), 138–149. Wang, JSY, Narasimhan, TN, Scholz, CH, 1988. Aperture correlation of a fractal frac-
Mohamad, AA., 2011. Lattice Boltzmann Method: Fundamentals and Engineering Ap- ture. J. Geophys. Res. 93 (B3), 2216–2224.
plications with Computer Codes. Springer Science and Business Media. Wang, L, Cardenas, MB, Slottke, DT, Ketcham, RA, Sharp, JM, 2015. Modification of
Mohammadipoor, OR, Niazmand, H, Mirbozorgi, SA, 2014. Alternative the local cubic law of fracture flow for weak inertia, tortuosity, and roughness.
curved-boundary treatment for the lattice Boltzmann method and its ap- Water Resour. Res. 51 (4), 2064–2080.
plication in simulation of flow and potential fields. Phys. Rev. E 89 (1), 013309. Witherspoon, PA, Wang, JSY, Iwai, K, Gale, JE, 1980. Validity of cubic law for fluid
Molz, FJ, Liu, HH, Szulga, J, 1997. Fractional Brownian motion and fractional Gaus- flow in a deformable rock fracture. Water Resour. Res. 16 (6), 1016–1024.
sian noise in subsurface hydrology: A review, presentation of fundamental prop- Wolf-Gladrow, D.A., 20 0 0. Lattice-Gas Cellular Automata and Lattice Boltzmann
erties, and extensions. Water Resour. Res. 33 (10), 2273–2286. Models: An Introduction. Springer Science and Business Media.
Montemagno, CD., Pyrak-Nolte, LJ., 1999. Fracture network versus single fractures: Xie, LZ, Gao, C, Ren, L, Li, CB, 2015. Numerical investigation of geometrical and hy-
measurement of fracture geometry with X-ray tomography. Phys. Chemistry draulic properties in a single rock fracture during shear displacement with the
Earth, Part A 24 (7), 575–579. Navier–Stokes equations. Environ. Earth Sci. 73 (11), 7061–7074.
Neuman, SP, 2005. Trends, prospects and challenges in quantifying flow and trans- Ye, ZY, Liu, HH, Jiang, QH, Zhou, CB, 2015. Two-phase flow properties of a horizontal
port through fractured rocks. Hydrogeol. J. 13 (1), 124–147. fracture: the effect of aperture distribution. Adv. Water Resour. 76, 43–54.
Newman, MS., Yin, X., 2013. Lattice Boltzmann simulation of non-Darcy flow in Zawada-Tomkiewicz, A., 2010. Estimation of surface roughness parameter based on
stochastically generated 2D porous media geometries. SPE J. 18 (01), 12–26. machined surface image. Metrol. Measure. Syst. 17 (3), 493–503.
Odling, NE., 1994. Natural fracture profiles, fractal dimension and joint roughness Zeng, Z, Grigg, R, 2006. A criterion for non-Darcy flow in porous media. Transport
coefficients. Rock Mech. Rock Eng. 27 (3), 135–153. Porous Media 63 (1), 57–69.
Ogilvie, SR., Isakov, E, Glover, PW., 2006. Fluid flow through rough fractures in rocks. Zhang, D, Papadikis, K, Gu, S, 2016. A lattice Boltzmann study on the impact of
II: a new matching model for rough rock fractures. Earth Planet. Sci. Lett. 241 the geometrical properties of porous media on the steady state relative perme-
(3), 454–465. abilities on two-phase immiscible flows. Adv. Water Resour. http://dx.doi.org/10.
Pan, C, Luo, LS, Miller, CT, 2006. An evaluation of lattice Boltzmann schemes for 1016/j.advwatres.2015.12.015. http://dx.doi.org/.
porous medium flow simulation. Comput. Fluids 35 (8), 898–909. Zhang, X, Crawford, JW, Young, IM, 2008. Does pore water velocity affect the reac-
Qian, YH., d’Humières, D, Lallemand, P., 1992. Lattice BGK models for Navier-Stokes tion rates of adsorptive solute transport in soils? Demonstration with pore-scale
equation. Europhys. Lett. 17 (6), 479. modelling. Adv. Water Resour. 31 (3), 425–437.
Sahimi, M., 2011. Flow and transport in porous media and fractured rock: from clas- Zhou, JQ, Hu, SH, Fang, S, Chen, YF, Zhou, CB, 2015. Nonlinear flow behavior at low
sical methods to modern approaches. John Wiley and Sons. Reynolds numbers through rough-walled fractures subjected to normal com-
Saravanathiiban, DS, Kutay, ME, Khire, MV, 2014. Effect of macropore tortuosity and pressive loading. Int. J. Rock Mech. Mining Sci. 80, 202–218.
morphology on preferential flow through saturated soil: a lattice Boltzmann Zimmerman, RW, Al-Yaarubi, A, Pain, CC, Grattoni, CA, 2004. Nonlinear regimes of
study. Comput. Geotech. 59, 44–53. fluid flow in rock fractures. Int. J. Rock Mech. Mining Sci. 41, 1A27.
Scesi, L, Gattinoni, P, 2007. Roughness control on hydraulic conductivity in fractured Zimmerman, RW, Bodvarsson, GS, 1996. Hydraulic conductivity of rock fractures.
rocks. Hydrogeol. J. 15 (2), 201–211. Transport Porous Media 23 (1), 1–30.
Schmittbuhl, J, Steyer, A, Jouniaux, L, Toussaint, R, 2008. Fracture morphology and Zou, L, Jing, L, Cvetkovic, V, 2015. Roughness decomposition and nonlinear fluid flow
viscous transport. Int. J. Rock Mech. Mining Sci. 45 (3), 422–430. in a single rock fracture. Int. J. Rock Mech. Mining Sci. 75, 102–118.