See discussions, stats, and author profiles for this publication at: [Link]

net/publication/273023734

Mean and Variance of Mobilized Shear Strength for Spatially Variable Soils
under Uniform Stress States

Article in Journal of Engineering Mechanics · March 2014

DOI: 10.1061/(ASCE)EM.1943-7889.0000667

CITATIONS READS

45 231

3 authors, including:

Jianye Ching Kok-Kwang Phoon

National Taiwan University Singapore University of Technology and Design
271 PUBLICATIONS 8,704 CITATIONS 518 PUBLICATIONS 20,736 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Jianye Ching on 26 December 2015.

The user has requested enhancement of the downloaded file.

1 Mean and variance of the mobilized shear strength for spatially variable soils under uniform

2 stress states

4 Jianye Ching1 M. ASCE, Kok-Kwang Phoon2 F. ASCE, and Ping-Hsun Kao3

6 ABSTRACT

7 This study proposes a set of simple equations for the mean value and variance of the mobilized

8 shear strength for spatially variable soil masses subjected to uniform stress states. These equations

9 are fairly effective in explaining the complicated behaviors for the mobilized shear strengths,

10 regardless of stress states (e.g., compression or shear), spatial variability patterns (e.g., isotropic or

11 anisotropic), and inherent mean and variance of the random field. Two mechanisms that affect the

12 behaviors of the mobilized shear strength are identified: (a) the line averaging effect along the

13 potential slip curves, and (b) the emergent feature of a critical slip curve. The emergence is simply

14 associated with the slip curve with the minimum line averaged strength. In any realization of the

15 random field, it is not possible to know apriori the location of the minimum line average and hence,

16 it would not coincide with a prescribed line average. It is shown that the well known phenomenon

17 of critical scale of fluctuation is the result of the tradeoff between these two mechanisms.

18 Key words: random field; spatial variability; shear strength; scale of fluctuation; finite element

1
(Corresponding author) Professor, Dept of Civil Engineering, National Taiwan University, Taipei, Taiwan. Email:
jyching@[Link]. Phone: 886-2-33664328. Fax: 886-2-23631558.
2
Professor, Dept of Civil and Environmental Engineering, National University of Singapore, Singapore.
3
Graduate student, Dept of Civil Engineering, National Taiwan University, Taipei, Taiwan.
1
19 analysis

20

21 INTRODUCTION

22 Spatial variability of shear strength and other design parameters are commonly encountered in

23 naturally occurring materials such as soils and rocks. Engineers are fully cognizant of this spatial

24 variation, but there are strong practical reasons to idealize soil profiles as either homogeneous or in

25 the presence of significant geologic contrast, a handful of soil layers. It is important to distinguish

26 between the “local” shear strength and the “mobilized” shear strength. In this study, the local

27 strength refers to the spatially averaged undrained shear strength over a finite element. The

28 “point” undrained shear strength is assumed to vary spatially according to a realization of a random

29 field. The spatially averaged (local) undrained shear strength serves as an input parameter to the

30 constitutive equations governing a particular finite element. The mobilized strength refers to the

31 maximum stress that can be applied to the top of a rectangular soil specimen under plane strain

32 condition. This mobilized strength, denoted by fm, is the solution of a boundary value problem

33 obtained via finite element analysis in this study. In contrast to the local/point shear strength, it is

34 clear that fm depends on the trajectory of the critical slip curve passing through the spatially

35 varying soil mass and this trajectory is unknown a priori. The focus in this study is to relate this

36 “emergent” feature of the critical slip curve to the mean and variance of fm.

37 It is useful to briefly examine qualitatively if the emergent feature of the critical slip curve

38 could be ignored. In other words, could the position of a slip curve be prescribed and spatial
2
39 averaging (e.g., Vanmarcke 1977) be applied to determine the statistical properties of fm? A clear

40 distinction between taking the spatial average along an arbitrarily prescribed slip curve and along

41 the critical slip curve is immediately evident. The statistical properties of the former are purely a

42 function of the random field. The statistical properties of the latter must depend to some extent on

43 the physical problem, because the critical slip curve is not prescribed but is a solution of a boundary

44 value problem in a spatially variable medium as mentioned above. One should expect the

45 trajectory of the critical slip curve to be different for each realization of the random field. This

46 emergent feature makes it difficult to derive the statistical properties for fm of a spatially variable

47 soil mass. It is important to qualify that spatial averaging does not deal with these emergent

48 curves.

49 The objective of this study is to propose a set of closed-form equations for the mean and

50 variance for fm of a spatially variable soil mass that incorporates the emergent feature of the critical

51 slip curve explicitly. These equations are based on a key observation made by Ching and Phoon

52 (2012a) that fm is nearly the same as the minimum value among the averaged shear strengths along

53 the potential slip curves. The critical slip curve is one member of these potential slip curves.

54 These equations are able to mimic the rather complicated statistical behaviors of fm. In addition,

55 they are able to explain the phenomenon of the “critical scale of fluctuation” observed in literature,

56 which refers to the emergence of non-classical behaviors not found in homogeneous soils when the

57 scale of fluctuation (SOF) is comparable to some multiple of the characteristic length of the

58 structure (e.g., height of slope, diameter of tunnel, depth of excavation) (Fenton and Griffiths 2008).
3
59 This critical SOF is observed in other studies as well (Fenton and Griffiths 2003; Breysse et al.

60 2005; Soubra et al. 2008).

61

62 RANDOM FIELD AND SPATIAL AVERAGING

63 Spatial variabilities of soil properties are commonly modeled by random fields (Vanmarcke 1977).

64 Among random field models, stationary (or statistically homogeneous) random fields are widely

65 used due to their simplicity and possibly the only practical version that can be characterized

66 statistically from limited data (Phoon et al. 2003). A two dimensional stationary random field for

67 shear strength f(x,z) can be defined over a plane with x = horizontal coordinate and z = depth

68 coordinate. This random filed is characterized by three parameters: (a) inherent mean E(f), (b)

69 inherent variance Var(f), and (c) auto-correlation function. The auto-correlation function of a

70 stationary random field f(x,z) is defined as the correlation between two locations x apart

71 horizontally in the x-direction and z apart vertically in the z-direction:

CV(τ f (x,z),τ f (x+Δx,z+Δz))

72 ρ(Δx,Δz)=ρ(τ f (x,z),τ f (x+Δx,z+Δz))= (1)
Var(τ f (x,z))  Var(τ f (x+Δx,z+Δz))

73 where Var(.) denotes variance; CV(.,.) denotes covariance. The hypothesis of stationarity allows 

74 to be simplified as a function of x and z only, rather than the absolute coordinates (x, z) and

75 (x+x, z+z) [see left hand side of Eq. (1)]. The most popular auto-correlation model is the single

76 exponential (SExp) model:

77 ρ(Δx,Δz)=exp(- 2 Δx δx -2 Δz δz ) (2)

4
78 where x and z are respectively the SOFs in the x and z directions. Another popular

79 auto-correlation model is the squared exponential (QExp) model:

ρ(Δx,Δz)=exp -π  Δx δ x  -π  Δz δ z  
2 2
80   (3)

81 It is clear that the correlation decreases as x and z increase. Because of the manner in which

82 most natural soils are deposited, soil properties are expected to be strongly correlated over a small

83 interval and weakly correlated when measurement points are sufficiently far apart.

84 Vanmarcke (1977) pointed out that the spatial average of a soil property over a region D has a

85 mean value identical to the inherent mean but a variance less than the inherent one. Let the region

86 D be a rectangular domain defined by [x0-Lx/2 x0+Lx/2] and [z0-Lz/2 z0+Lz/2]. In other words, (x0,

87 z0) is the centre of a rectangle of horizontal length = Lx and vertical length = Lz. Mathematically,

88 the spatial average over D is defined as

x 0 +L x /2 z 0 +L z /2
1
89
D
τ =
f   τ f  x,z  dzdx (4)
Lx Lz x 0 -L x /2 z 0 -L z /2

90 The variance of fD is less than the inherent variance Var(f). To quantify the spatial averaging

91 effect, Vanmarcke (1977) further defined the variance reduction factor to be the variance of fD

92 divided by the inherent variance:

93 Γ2D = Var  τ fD  Var  τ f  (5)

94 For the SExp model given in Eq. (4), Vanmarcke (1977) showed that

95  
Γ2D =Γ2  n x   Γ2  n z  =  2n x -1+exp  -2n x   2n 2x   2n z -1+exp  -2n z   2n 2z  (6)

96 where nx = Lx/x and nz = Lz/z, i.e., nx and nz are the equivalent numbers of SOFs in the x and z

5
97 directions, respectively. Note that D 2 is a decreasing function of nx and nz.

98

99 MECHANISMS OF MOBILIZED SHEAR STRENGTH

100 A key observation made by Ching and Phoon (2012) was that fm is nearly the same as the

101 minimum value among the line averages along the potential slip curves, regardless of stress state

102 and spatial variability pattern. The significance of this statement is explained in detail in Figure 1.

103 Figure 1a shows a realization of the f(x,z) random field on the Lx  Lz domain shown in Figure 2,

104 where the sections A, B, and C in Figure 1a represent the locations of three “potential” slip curves

105 (zA, zB, and zC in Figure 1a characterize the vertical positions of these three sections). There are

106 infinite potential slip curves, but for clarity only three of them are shown. These potential slip

107 curves are parallel and it is observed in Ching and Phoon (2012) that their inclination angles are

108 largely dominated by mechanics rather than by the spatial variability [inclination angle  

109 45o+(friction angle)/2 if the principle stress is vertical (e.g., compression test) and   (friction

110 angle)/2 if the principle stress inclines at 45o (e.g., pure shear test)]. Mollon et al. (2011) observed

111 that the failure mechanism at a pressurized tunnel face in spatially varying sand is largely

112 dominated by mechanics as well. Only one of the potential slip curves is the critical slip curve.

113 The position of the critical slip curve, say vertical coordinates of the left end of the curve, is quite

114 random and depends on the random field realizations.

115 The sectioned f profiles along the three potential slip curves are plotted in Figure 1b. The

116 average of each profile (the dashed lines in Figure 1b) is denoted by fLA and will be referred as the
6
117 “line average” for each section. Spatial averaging along each section (A, B, or C) produces a

118 single line average value fLA. Because the elevation of these three sections be zA, zB, and zC, their

119 line averages are therefore denoted by fLA(zA), fLA(zB), and fLA(zC). There are an infinite

120 number of potential slip curves. Each potential slip curve has a line average fLA, and the resulting

121 continuous line average process fLA(z) forms a one-dimensional (1-D) random field. This 1-D

122 random field is illustrated in Figure 1c. It is not the same as the random field for the shear strength

123 at a point, f(x,z). The line average process fLA(z) is a stationary 1-D Gaussian random field

124 defined on the interval [0, Lz – Lxtan()] with mean equal to E(fLA), variance equal to Var(fLA),

125 and SOF equal to LA. The size of the interval is Lz – Lxtan() because it is the vertical distance

126 between the topmost and bottom-most potential slip curves, as seen in the right plot of Figure 2.

127 Because the critical slip curve has the minimum line average among all potential slip curves, fm is

128 simply the minimum value of this 1-D random field:

129 τ fm = min τ fLA  z  (7)

z

130 In summary, it is useful to keep three distinct random fields and their respective means,

131 variances, and scales of fluctuations (SOFs) in mind:

132 1. f(x,z)  2-D Gaussian random field indexed on the domain Lx  Lz that describes the spatial

133 variation of the “point” shear strength. The mean is E(f) and the variance is Var(f). The

134 SOFs are x and z in the x and z directions, respectively.

135 2. Sectioned f(x,z)  1-D Gaussian random field indexed along a prescribed potential slip

136 curve of inclination angle  and length Lp = Lx/cos() (see the right plot in Figure 2) that
7
137 describes the spatial variation of the point shear strength along the slip curve. The mean is

138 E(f), the variance is Var(f), and the SOF is p. p is the SOF along the direction of the

139 potential slip curve (not equal to x and z). We further define np = Lp/p is the equivalent

140 number of SOF along the potential slip curve. The notion of np is illustrated in Figure 1b

141 (Section A) – it basically counts how many segments of size p are within the entire

142 potential slip curve. For the particular example shown in Figure 1b (Section A), Lp =

143 1.28p, i.e., np = 1.28. In the Appendix, the calculation steps for np for a general case with

144 potential slip curve of arbitrary inclination angle and with arbitrary x and z are presented.

145 3. fLA(z)  1-D Gaussian random field indexed on the vertical depth interval [0, Lz – Lxtan()]

146 that describes the spatial variability of the line averaged shear strength. The mean is

147 E(fLA), the variance is Var(fLA), and SOF is LA. Note that LA is not the same as p.

148 The former parameter describes how the line average fLA(z) varies with depth, while the

149 latter parameter describes how the point shear strength f(x,z) varies along the potential slip

150 curve. We further define nLA = [Lz – Lxtan()]/LA to be the equivalent number of LA

151 within fLA(z). This parameter controls the emergent feature of a critical slip curve, which

152 results from the minimization process fm = minzfLA(z). The notion of nLA is illustrated in

153 Figure 1c – it basically counts how many segments of length LA are within the entire

154 interval [0, Lz – Lxtan()]. A large nLA implies that there are more fluctuation cycles in

155 fLA(z). For the particular example in Figure 1c, Lz – Lxtan() = 1.29  LA, i.e., nLA = 1.29.

156 In the Appendix, the calculation steps for nLA for a general case with potential slip curve of
8
157 arbitrary inclination angle  and with arbitrary x and z are presented.

158 Finally, the minimum value of fLA(z) is denoted by fm. The value and location of fm depend on

159 the realization of f(x,z).

160 Complicated behaviors of the mobilized shear strength

161 The behaviors of fm are fairly complicated and depend on spatial variability patterns and stress

162 states. To illustrate these complicated behaviors, a series of random field finite element analyses

163 (FEA) is carried out using the Mohr Coulomb (c, ) strength model. A rectangular plane strain

164 domain, Lx  Lz = 12.8 m  48 m, is adopted as shown in the left plot in Figure 2. The size of each

165 finite element is 0.1 m  0.1 m. Hence, this rectangular domain D is discretized into 61,440

166 elements. The bottom boundary is supported on rollers, and the lower-left-most node is a hinge to

167 prevent rigid body translation in the x direction. The unit weights of all elements are zeros to

168 ensure uniform vertical stress, the Young’s modulus is prescribed a high value to expedite

169 convergence (400 MN/m2), and the Poisson ratio is 0.3. The spatially variable shear strength c = f

170 at a point is simulated by stationary Gaussian random fields with inherent mean E(f) = 50 kN/m2,

171 inherent standard deviation Var(f)0.5 = 10 kN/m2, horizontal SOF = x, and vertical SOF = z.

172 Element-level averaging within each 0.1 m  0.1 m element is adopted to generate a discretized f

173 value for each finite element. Measures are undertaken where appropriate to ensure that excessive

174 element-level averaging reported in Ching and Phoon (2012) is mitigated. To keep the problem

175 simple, the shear strength f is assumed to be independent of confining pressure, i.e.,  = 0o.

176 Two types of uniform stress states are considered for the time being: compression and pure
9
177 shear tests. For the compression test, the two lateral boundaries are free. An axial compression

178 stress is applied at the top boundary until FEA fails to converge. Undesirable shear traction will

179 not be developed on the bottom boundary because the boundary is supported on rollers. The

180 stress-strain curve (see Figure 3a) is then plotted, and the yield axial stress applied on the top

181 boundary (y) is identified. Note that the analysis in this paper is limited to the  = 0o condition,

182 under which f = c = the radius of the Mohr circle at failure. As a result, fm = the radius of the

183 Mohr circle at failure = y/2. This FEA simulation is similar to the unconfined compression (UC)

184 test in laboratory, except that it is in a plane strain condition, rather than a triaxial condition. The

185 UC test is adopted here because the initial confining pressure has no effect on fm (due to  = 0o

186 hypothesis). For the pure shear test, shear stress of equal magnitude is applied at the four

187 boundaries until FEA fails to converge. The stress-strain (shear stress versus shear strain) curve

188 (see Figure 3b) is then plotted, and the yield shear stress y is identified. Under the  = 0o

189 condition, fm = the radius of the Mohr circle at failure = y.

190 Two types of spatial variability are considered: (a) x = z =  (isotropic random field) and (b)

191 x = , z =  (anisotropic random field). Together with the two types of stress states

192 (compression and pure shear), there are four possible scenarios: (a) Scenario CI – compression test

193 with x = z; (b) Scenario CA – compression test with x = ; (c) Scenario SI – pure shear test with

194 x = z; and (d) Scenario SA – pure shear test with x = . The first letter denotes the stress state

195 (C: compression, S: pure shear), while the second letter denotes the random field type (I: isotropic

196 x = z, A: anisotropic x = ). For the time being, the SExp model [Eq. (2)] is adopted to produce
10
197 the detailed correlation structure of the random field. The effect of a different auto-correlation

198 model will be presented in a later section. For each scenario, 100 random samples of fm are

199 generated for the following 5 chosen SOFs:  = 0.1 m, 1 m, 10 m, 100m, and 1000 m. The case

200 with  = 1000 m is close to a homogeneous case.

201 Figure 4 shows the normalized sample means and normalized sample variances of the 100

202 random samples of fm at various SOFs for all four scenarios. Note that the  shown in this figure

203 is z for SA and CA and is x = z for SI and CI. The behaviors of the sample means and variances

204 are fairly complicated. The phenomenon of the critical SOF reported in Fenton and Griffiths

205 (2003), Breysse et al. (2005), and Soubra et al. (2008) is observed for scenarios CI, SI, and CA – the

206 sample means are minimal at  = 10 m. This critical  is roughly equal to the width of the soil

207 specimen, which is also roughly equal to the length of the slip curve. However, a critical  does

208 not exist for SA. The sample variances vary over a wide range depending on the stress state (C or

209 S) and spatial variability pattern (I or A). The means and variances of the spatial averages over the

210 entire domain D are also plotted as solid curves for comparison. It is clear that the sample means

211 of fm are always less than those predicted by spatial averaging, while the sample variances of fm

212 are always larger than those predicted by spatial averaging. Although the behaviors of fm are

213 fairly complicated, it will be demonstrated later that the proposed equations [Eq. (13) in a later

214 section] can effectively reproduce such complicated behaviors regardless of stress state and spatial

215 variability pattern.

216
11
217 STATISTICAL PROPERTIES FOR MOBILIZED SHEAR STRENGTH

218 Model for E(fm)

219 The following equation is proposed to predict the mean value of the mobilized shear strength:

220 E  τ fLA  -E  τ fm  =C  n LA   std  τ fLA  (8)

221 where C(nLA) is an increasing function of nLA that satisfies C(0) = 0, and std(fLA) = Var(fLA)0.5 is

222 the standard deviation of fLA(z). The rationale of this equation is simple. First of all, because

223 fm = minzfLA(z), it is clear that E(fm) must be less than E(fLA), and E(fLA)–E(fm) is the positive

224 difference between the point mean of the line average process fLA(z) and the mean value of the

225 mobilized shear strength. This difference should be proportional to std(fLA) – a large variability

226 in fLA(z) will amplify this difference. This can be seen from the left plot of Figure 5 that shows

227 the effect of std(fLA) – the larger std(fLA), the larger the difference in E(fLA)–E(fm). This

228 difference should also increase with increasing nLA – more fluctuations in the fLA(z) profile will

229 also amplify this difference. This can be seen from the right plot of Figure 5 that shows the effect

230 of LA ( 1/nLA) – the larger nLA, the larger the difference in E(fLA)–E(fm). Moreover, when nLA

231 = 0 [no fluctuation in fLA(z)], the difference should be zero, i.e., C(0) = 0. The C(nLA) function

232 will be calibrated later using simulated data.

233 Model for Var(fm)

234 The following equation is proposed to predict the variance of the mobilized shear strength:

235 Var  τ fm  =Var  τ fLA   Ω 2  n LA  =Var  τ f   Γ 2  n p   Ω 2  n LA  (9)

236 where 2(nLA)  1 is the variance reduction factor due to the minimization process fm = minzfLA(z).
12
237 This is based on the property of order statistics that the variance of min(X1, X2, …, Xn) (X1, X2, …,

238 Xn are independent identically distributed) is less than the variance of Xi. In the extreme case

239 where nLA = 0, there is no fluctuation in fLA(z), and Var(fm) will be identical to Var(fLA). On the

240 other hand, if nLA is large, there are many fluctuations in fLA(z), and Var(fm) will be less than

241 Var(fLA). As a result, the behavior of 2 is similar to that of 2: it is a decreasing function of nLA

242 and 2(0) = 1. However, 2 decays with nLA less rapidly than 2. This 2(nLA) function will be

243 calibrated later using simulated data.

244 Mean and variance of the 1-D random field fLA(z)

245 Because fLA(z) is the result of a 1-D line averaging of the point process f(x, z) along a potential

246 slip curve, it is normally distributed with E(fLA) = point mean E(f) and Var(fLA) = point variance

247 Var(f) multiplied by the variance reduction factor 2 (Vanmarcke 1977):

248 Var  τfLA  =Var  τf   Γ2  n p  (10)

249 where the variance reduction factor 2(np) characterizes the line averaging effect along the potential

250 slip curve

  2n p -1+exp  -2n p    2n 2p  for SExp model

  
Γ2  n p  = 
251
  
  πn p  erf π×n p +exp  -πn p  -1  πn p  for QExp model
2 2 (11)

252 and erf(.) is the error function

u
2
253 erf  u  = e
-t 2
dt (12)
π 0

254 Main equations

255 Combining Eqs. (8) and (9) gives the following equation:
13
256 E  τ fm  =E  τ f  -C  n LA   Var  τ f  ×Γ 2  n p  Var  τ fm  =Var  τ f   Γ 2  n p   Ω 2  n LA  (13)

257 In the above, we have included the observation that E(fLA) = E(f). Compared to the equations for

258 the spatial average fD,

259 E  τ fD  =E  τ f  Var  τ fD  =Var  τ f   Γ 2  n x   Γ 2  n z  (14)

260 Eq. (13) shows three distinct features: (a) there is a discrepancy between E(f) and E(fm), (b) only

261 the averaging effect along the potential slip curve is relevant for fm; the spatial average over the

262 entire domain D is not relevant, and (c) part of the variance reduction for fm is due to the

263 minimization process fm = minzfLA(z) and is quantified by 2(nLA) in Eq. (13).

264 It is evident that there are two main mechanisms governing the statistics of fm: line averaging

265 along a potential slip curve and emergence of a critical slip curve. The line averaging mechanism

266 occurs on a potential slip curve and manifests itself in the term 2(np) in Eq. (13). This mechanism

267 is characterized by np – a large np will reduce the difference E(f)–E(fm) and will also reduce

268 Var(fm), and vice versa. The largest discrepancy between E(f) and E(fm) and the largest Var(fm)

269 occur when np = 0, i.e., when p = , when potential slip curves are parallel to a direction without

270 spatial variability. From an engineering perspective, conditions producing the smallest E(fm) and

271 the largest Var(fm) are adverse. Scenario SA is one such condition. In this condition, it could be

272 quite unconservative to use spatial average of in-situ tests or laboratory test results as the

273 characteristic value. The emergence mechanism underlying a critical slip curve manifests itself in

274 the two terms C(nLA) and 2(nLA) in Eq. (13), both related to the minimization process fm =

275 minzfLA(z). This mechanism is characterized by nLA – a large nLA will increase the difference
14
276 E(f) – E(fm) and will reduce Var(fm), and vice versa.

277 Critical scale of fluctuation

278 Equation (13) can explain why there is a critical  in Figure 4 for CI, CA, and SI but not for

279 SA. This is because for CI, CA, and SI, np and nLA both depend on . As an example, for SI, x

280 = z =  and  = 0o ( is the inclination angle of potential slip curves). It is clear that nLA  Lz/z =

281 48/, and np = Lx/x = 12.8/. Because C(nLA) is an increasing function of nLA, it is a decreasing

282 function of . In contrast, 2(np) is a decreasing function of np, so it is an increasing function of .

283 However, the product C(nLA)2(np)0.5 is neither a completely increasing nor completely

284 decreasingly function of  – there is a critical  where C(nLA)2(np)0.5 is maximized – or where

285 E[fm] is minimized. This phenomenon is closely related to the critical SOF observed by Fenton

286 and Griffiths (2003), Breysse et al. (2005), and Soubra et al. (2008). This critical  does not exist

287 for SA because np does not depend on . Recall that for SA, x =  and  = 0o, so np = Lx/x = 0

288 and 2(np) = 1. Hence, C(nLA)2(np)0.5 = C(nLA), and it always decreases with .

289

290 CALIBRATION OF C(nLA) and 2(nLA)

291 The previous section proposes Eq. (13) for the mean and variance of fm. In this section, simulated

292 data from CI, CA, SI, and SA will be used to calibrate the C(nLA) and 2(nLA) functions in this

293 equation. As mentioned earlier, fm samples for four scenarios (CI, CA, SI, and SA) and five SOFs

294 ( = 0.1 m, 1 m, 10 m, 100m, and 1000 m) are simulated – there are twenty combinations in total.

295 For each combination, 100 fm samples are simulated, and these 100 samples can be used to obtain
15
296 the sample mean E(fm) and sample variance Var(fm) for each combination. By following the

297 steps in the Appendix, nLA and np can be determined for each combination. Note that nLA and np

298 are not parameters required by FEA simulations – FEA simulations only need the realizations of

299 f(x, z). Once f(x, z) is given to the FEA mesh, The FEA will automatically seek out the critical

300 slip curve among all the potential slip curves. A sample value of C(nLA) can then be obtained for

301 each combination as [E(f)–E(fm)]/[Var(f)2(np)]0.5, and a sample value of 2(nLA) can then be

302 obtained as Var(fm)]/[Var(f)2(np)]. Note that E(f) = 50 kN/m2 and Var(f) = 102 (kN/m2)2 are

303 known values, and 2(np) can be easily calculated based on Eq. (8). At the end, there will be

304 twenty data points for the nLA-C(nLA) relation and another twenty data points for the nLA-2(nLA)

305 relation. These data points are shown in the upper two plots in Figure 6. Rough unique trends

306 are observed, and they are

3.4  -2.0 
307 C  n LA  = Ω 2  n LA  =exp  -0.8  (15)
1+2.0  n -0.6
LA  1+3.0  n LA 

308 Note that they satisfy C(0) = 0 and 2(0) = 1. The scatter with respect to the trends is in part due

309 to the small sample size = 100 for estimating sample means and sample variances. Compared to

310 the variance reduction 2 in Eq. (8), the variance reduction factor 2(nLA) decays with nLA fairly

311 slowly (see the right plots in Figure 6). This is because their mechanisms are different: 2 in Eq.

312 (8) is due to spatial (or line) averaging, but 2 is due to the minimization fm = minzfLA(z).

313 Note that the two trends are calibrated by data points from CI, CA, SI, and SA only.

314 Moreover, the inherent mean E(f) and inherent variance Var(f) of the calibration cases are fixed at

16
315 50 kN/m2 and 102 (kN/m2)2, respectively. Nonetheless, further verifications show that these two

316 calibrated trends work equally well for scenarios with different stress states (other than compression

317 ‘C’ and shear ‘S’), combinations of x and z, inherent mean, and inherent variance. Note that the

318 autocorrelation function is single exponential in all the scenarios. Figure 6 shows the data points

319 of C(nLA) and 2(nLA) for these extra verification cases. They follow the calibrated trends

320 reasonably well. Some scatter exists showing that C(nLA) and 2(nLA) are not perfectly unique

321 functions.

322 To demonstrate how well Eq. (13) with the calibrated C(nLA) and 2(nLA) in Eq. (15) can be

323 used to mimic the complicated behaviors of E(fm) and Var(fm), Figure 7 shows the simulated

324 results for CI and SA. The behaviors in E(fm) and Var(fm) for these two scenarios are quite

325 different, especially in E(fm) – SA does not show a critical , but CI does. Nonetheless, Eq. (13)

326 is able to give E(fm) and Var(fm) values that are fairly close to the actual data for these two

327 drastically different scenarios. The predicted E(fD) and Var(fD) based on spatial averaging are

328 also plotted for comparison – they do not agree with the simulated data.

329 To demonstrate how well Eq. (13) with the calibrated C(nLA) and 2(nLA) in Eq. (15) can be

330 used to mimic the phenomenon of the critical , Figure 8 shows the sample means E(fm) obtained

331 for CI, CA, SI, and SA. Note that these data points were used to calibrate C(nLA) and 2(nLA). It

332 is clear that the phenomenon of the critical  is successfully captured by Eq. (13). As discussed

333 earlier, this phenomenon is due to the tradeoff between the two terms C(nLA) and 2(np), i.e., the

334 tradeoff between the line averaging effect and the emergent critical slip curve. For CI, CA, and SI,
17
335 the former increases with nLA and the latter decreases with nLA. For SA, such tradeoff does not

336 exist because 2(np) = 1 regardless of nLA.

337 QExp model

338 Equation (15) and Figure 6 show the calibration results of C(nLA) and 2(nLA) for the SExp model.

339 For the QExp model, Eq. (13) still holds true, but with a subtle difference from Eq. (8) that the

340 variance reduction factor 2(np) for the QExp model is different from that for the SExp model [see

341 Eq. (11)]. Moreover, the similar calibration shows that Eq. (15) should be replaced by the

342 following equations:

3.45  -2.0 
343 C  n LA  = Ω 2  n LA  =exp  -0.8  (16)
1+3.3  n -0.9
LA  1+7.0  n LA 

344 Figure 9 shows the calibration data points together with the calibrated C(nLA) and 2(nLA) (solid

345 lines). Note that the calibrated C(nLA) and 2(nLA) trends for the QExp model somewhat deviate

346 from those for the SExp model (dashed lines).

347

348 CONCLUDING REMARKS

349 This study proposes simple equations for the mean and variance of the mobilized shear strength of a

350 spatially variable soil mass subjected to uniform stress state. The resulting equations [Eqs. (13)

351 and (15)] are fairly simple but are able to effectively predict the mean value and variance of the

352 mobilized shear strength. There are two key input parameters (nLA, np). The former parameter

353 nLA is the equivalent number of scale of fluctuation (SOF) in a one-dimensional random field

18
354 resulting from line averaging along the potential slip curves. This parameter characterizes the

355 emergent feature of a critical slip curve that is rarely explored in literature. The latter parameter np

356 is the equivalent number of SOF along the potential slip curves and characterizes the line averaging

357 effect along the potential slip curves. It is not well emphasized in the literature that the spatial

358 variation along a potential slip curve is not the same as the ambient spatial variability across the

359 entire soil mass. It is evident that these two mechanisms (emergent critical slip curve and line

360 averaging along potential slip curves) govern the mobilized shear strength. In fact, we believe that

361 this is the first occasion in which the phenomenon of a critical SOF observed in numerous random

362 finite element analyses has been explained as a tradeoff between these two mechanisms.

363 The equations proposed by Vanmarcke (1977) are applicable to the spatially averaged shear

364 strength over a prescribed domain. This was indeed the intention of work presented in Vanmarcke

365 (1977). However, this classical concept of spatial averaging can deviate from the mobilized shear

366 strength in significant ways.

367 There are limitations for the results of this study. First, the proposed equations have been

368 validated for soil masses subjected to uniform stress states thus far. This is a rather idealized

369 situation, but it is a necessary first step to examine a fairly fundamental and complex issue relating

370 to failure in a spatially variable medium. These equations should not be applied to problems with

371 more complex stress states at present. Second, the proposed equations are for soil masses where

372 potential slip curves are free to evolve. For instance, for retaining wall problems the slip curve

373 typically passes through the toe of the wall, the proposed equations may not be applicable. In
19
374 other words, we have not studied the interaction of a structure with a spatially variable soil mass.

375 Furthermore, it is not certain whether the equations are applicable to three dimensional problems

376 subjected to uniform stress states and with un-constrained slip curves. Nonetheless, for general

377 problems with more complex stress states and with slip curves constrained by structural interaction,

378 the basic ideas and mechanisms of the equations might still hold true.

379

380 ACKNOWLEDGEMENTS

381 This study was under the sponsorship of the National Science Council (NSC) of Republic of China,

382 under the project NSC100-2628-E-002-040-MY3. The first author would like to acknowledge

383 such a gracious support from NSC.

384

385 REFERENCES

386 Breysse, D., Niandou, H., Elachachi, S., and Houy, L. (2005). A generic approach to soil-structure

387 interaction considering the effects of soil heterogeneity, Geotechnique, 55(2), 143-150.

388 Ching, J. and Phoon, K.K. (2012). Mobilized shear strength of spatially variable soils under simple

389 stress states, conditionally accepted by Structural Safety.

390 Fenton, G.A. and Griffiths, D.V. (2003). Bearing capacity prediction of spatially random c-φ soils.

391 Canadian Geotechnical Journal, 40, 54-65.

392 Fenton, G.A. and Griffiths, D.V. (2008). Risk Assessment in Geotechnical Engineering. John Wiley

393 & Sons, New York.

20
394 Jha, S.K. and Ching, J. (2012). Simulating spatial averages of stationary random field using Fourier

395 series method, to appear in ASCE Journal of Engineering Mechanics.

396 Mollon, G., Phoon, K.K., Dias, D., and Soubra, A.-H. (2011). Validation of a new 2D failure

397 mechanism for the stability analysis of a pressurized tunnel face in spatially varying sand.

398 ASCE Journal of Engineering Mechanics, 137(1), 8-21.

399 Phoon, K.K., Quek, S.T., and An, P. (2003). Identification of statistically homogeneous soil layers

400 using modified Bartlett statistics. ASCE Journal of Geotechnical and Geoenvironmental

401 Engineering, 129(7), 649-659.

402 Soubra, A.H., Youssef Abdel Massih, D.S., and Kalfa, M. (2008). Bearing capacity of foundations

403 resting on a spatially random soil. GeoCongress 2008: Geosustainability and Geohazard

404 Mitigation, ASCE Geotechnical Special Publication No. 178, 66–73.

405 Vanmarcke, E.H. (1977). Probabilistic modeling of soil profiles. ASCE Journal of Geotechnical

406 Engineering Division, 103(11), 1227-1246.

407

408 APPENDIX CALCULATION STEPS FOR nLA AND np

409 Consider a general random field with x  z and the principle stress inclines at an angle of 

410 clockwise from the vertical line. There are two sets of potential slip curves – one set inclines at an

411 angle of 1 = 45o+/2- counter-clockwise from the horizontal line, and the other inclines at an

412 angle of 2 = 45o+/2+ clockwise from the horizontal line.

413 The first part of this appendix shows how the equivalent number of SOF (nLA) of fLA(z) can be
21
414 estimated for any potential slip curve that inclines at an angle of  from the horizontal line. The

415 second part of this appendix shows how np = Lp/p can be determined.

416 Determination of nLA

417 Let us consider two potential slip curves (P1 and P2 shown in the right plot of Figure 2) with vertical

418 spacing of S, both with inclination angle of . Let us further denote the line averages along these

419 two curves by f LA1 and f LA2:

1
τ LA1
f =  τf  x L1 +t1  x R1 -x L1  ,z L1 +t1  z R1 -z L1   dt1
0
420 1 (17)
τfLA2 =  τf  x L2 +t 2  x R2 -x L2  ,z L2 +t 2  z R2 -z L2   dt 2
0

421 where (xL1, zL1) and (xL2, zL2) are the left-end nodes of P1 and P2, and (xR1, zR1) and (xR2, zR2) are

422 the right-end nodes; t1 and t2 are local coordinates on P1 and P2 (t = 0 is the left-end node, and t = 1

423 is the right-end node). Note that xL1 = xL2, xR1 = xR2, zL2 - zL1 = zR2 - zR1 = S, xR1 - xL1 = xR2 - xL2

424 = Lx, zR1 - zL1 = zR2 - zL2 = Lxtan(). The correlation coefficient between fLA1 and fLA2 is

1 1

425 ρτ LA1

f ,τ LA2
f  =ρ S =   ρ Δx=L  t -t  ,Δz=L tan β  t -t  -S dt dt
x 1 2 x 1 2 2 1 Γ2  n p  (18)
0 0

426 where (fLA1,fLA2) is the correlation coefficient between fLA1 and fLA2; the correlation (x,z)

427 in the integrant is the correlation between two point processes of f with horizontal distance x and

428 vertical distance z; np = Lp/p is the equivalent number of SOF along the potential slip curves.

429 The notation (S) is used to denote the correlation between fLA1 and fLA2 which are apart with a

430 vertical spacing “S”. The scale of fluctuation for the line averaging process fLA(z), denoted by

431 LA, is therefore

22
 

432 δLA =  ρ  S dS (19)

-

433 The equivalent number of SOF in fLA(z) is therefore

434 n LA = Lz -Lx tan  β   δLA (20)

435 where Lz – Lxtan() is the vertical spacing between the topmost and bottom-most slip curves (see

436 the right plot of Figure 2). The derivations for LA are different for the SExp and QExp models,

437 which are given below.

438 SExp model

439 For the SExp model, the integration in Eq. (19) has the following analytical solution:

 1 1

   exp - 2L x δ x t1 -t 2 - 2L x tan  β  δ z t1 -t 2 -S  L x tan  β    dt 2 dt1dS

δLA = -  0 0
 2n p +exp  -2n p  -1 2n p2
 
 r1 +exp(-r1 )-1 r12  r3  δ z +  r2 +exp(-r2 )-1 r22  r3  δ z +  r1 +exp(-r1 )-1 r12  r4  δ z 
2 
+  r5 -r4  r2  δ z +  r4 -r5  r2  δz +  r1 -1 r5 +exp(-r1 )  r6  r1  δ z - r7 r1  δ x + r7 r2  δ x 
2 2

=
 2n p +exp  -2n p  -1 2n p2
440   (21)
where
V=Lx tan  β  n x = Lx δx n v = V δz r1 =2n x +2n v r2 =2n x -2n v r3 = exp  -2n v  2
r4 = 1-exp  -2n v  2 r5 = 1-exp  -2n x    2tan  β  
r6 = exp  2n v  -1 2 
r7 = tan  β  2  1-exp  -2n x   2n x -exp  -2n x  
441 QExp model

442 For the QExp model, the integration in Eq. (19) cannot be obtained analytically:

23
  1 1

   exp -πL  t -t  δ 2x - πL2x tan  β   t1 -t 2 -S Htan  β   δz2  dt 2 dt1dS

2 2 2 2
x 1 2

δ LA = - 0 0
 πn p  erf
  
π  n p +exp  -πn 2p  -1 πn 2p

1 1
443 (22)
δ  exp - πL2x  t1 -t 2  δ2x  dt1dt 2
2
z  
0 0
=
 πn p  erf
  
π  n p +exp  -πn 2p  -1 πn p2


444 where erf(.) is the error function. Numerical integration can be taken to evaluate the integral.

445 There are two sets of potential slip curves with inclination angles 1 and 2. Let the

446 corresponding SOF count numbers of fLA be n1 and n2, respectively. Even though any two

447 potential slip curves belonging to set 1 may be considered independent, a potential slip curve in set

448 1 may not be independent of another slip curve in set 2. Hence, the final estimation for n should

449 not be n1 + n2. Because fm for the set with a larger n value is likely to be less than that for the

450 other set, the final estimation for n is taken to be max(n1, n2).

451 Determination of np

452 SExp model

453 For two points on the potential slip curve that are s apart, the correlation between them is

454 exp(-2/p|s|), where p is the SOF along the potential slip curve. Since these two points are

455 |s|cos() apart in the horizontal direction and |s|sin() apart in the vertical, their correlation has

456 another expression: exp[-2/x|s|cos()]  exp[-2/z|s|sin()]. It is then clear that

457 exp(-2/p|s|) = exp[-2/x|s|cos()]  exp[-2/z|s|sin()], and hence p = 1/[cos()/x+

458 sin()/z]. Furthermore, Lp = 12.8/cos(). Therefore, np = Lp/p = 12.8[1/x+ tan()/z].

459 QExp model

24
460 For the same two points, the correlation between them is exp(-/p2s2). Another expression

461 for the correlation is exp[-/x2s2cos2()]  exp[-/z2s2sin2()]. It is then clear that

462 exp(-/p2s2) = exp[-/x2s2cos2()]  exp[-/z2s2sin2()], and hence p =

463 1/[cos2()/x2+sin2()/z2]0.5. Furthermore, Lp = 12.8/cos(). Therefore, np = Lp/p =

464 12.8[1/x2+tan2()/z2]0.5.

465 There are two sets of potential slip curves with inclination angles 1 and 2. Let the

466 corresponding SOF count numbers of fLA be n1 and n2, respectively. The np value should be

467 determined based on the set with a larger nLA.

468 Figure A-1 shows the variations of nLA and np with respect to  for CI, CA, SI, and SA under

469 the SExp model, while Figure A-2 shows the variations under the QExp model. The (1, 2) are

470 (45o, -45o) for CI and CA and are (0o, 90o) for SI and SA. The (x, z) are equal for CI and SI and

471 are unequal (x = ) for CA and SA. It is clear that nLA increases with decreasing . The nLA

472 values for CA are particularly small. This is because x =  (horizontal layers) and the potential

473 slip curves incline at an angle of 45o for scenario CA. As a result, any two potential slip curves

474 with vertical distance less than 12.8 m share common layers and are hence dependent, regardless of

475 the value of z. In fact, nLA seems to have an upper bound for CA. Note that np = 0 for SA

476 regardless of .

477

478

25
479 SYMBOL LIST

480  = friction angle

481  = inclination angle of a potential slip curve
482  = scale of fluctuation
483 (.,.) = auto-correlation function
484 2 = variance reduction factor
485 D2 = variance reduction factor for the two-dimensional spatial averaging over region D
486 f = shear strength
487 f(x,z) = point process of the f random field
488 fD = spatial average of the f(x,z) random field over region D
489 fLA = line average of f along a potential slip curve
490 fLA(z) = the line average process of fLA (a one-dimensional random field)
491 fm = mobilized shear strength
492 LA = scale of fluctuation of the 1-D random field for the line average process fLA(z)
493 p = scale of fluctuation along the potential slip curve
494 x = horizontal scale of fluctuation
495 y, y = yield stresses in the stress-strain curve
496 z = vertical scale of fluctuation
497 c = cohesion
498 C(nLA) and 2(nLA): two functions capturing the emergent feature of the critical slip curve
499 CV(.,.) = covariance
500 D = rectangular domain
501 E(f) = inherent (point) mean of the f(x,z) random field
502 E(.) = mean value
503 Lp = length of the potential slip curve
504 Lx and Lz = width and height of the region D
505 nLA = [Lz – Lxtan()]/LA = the equivalent number of SOF within fLA(z)
506 np = Lp/p = the equivalent number of SOF along the potential slip curve
507 nx = Lx/x = the equivalent number of SOF in the x direction
508 nz = Lz/z = the equivalent number of SOF in the z direction
509 std(.) = standard deviation
510 Var(f) = inherent (point) variance of the f(x,z) random field
511 Var(.) = variance
512 x = horizontal coordinate
513 z = vertical coordinate
514

26
515

516

517 Figure 1a Figure 1b Figure 1c

518 Figure 1 (Figure 1a) a realization of the f(x,z) random field and the locations of sections A, B,
519 and C; (Figure 1b) the sectioned f(x,z) profiles and the line averages fLA; (Figure 1c) the
520 continuous line average process fLA(z).

521

27
522
523
524 Figure 2 (left) finite element model for the plane strain soil specimen; (right) potential slip curves.
525
526
527
528
529

530
531
532 Figure 3 Stress-strain curves for (left) compression test and (right) shear test.

533

28
534
535 Figure 4 Normalized sample means and sample variances of fm for all four scenarios. Upper
536 two plots are for CI and SI, while the lower two plots are for CA and SA.
537

29
538

539 Figure 5 (left plot) effect of std[fLA(z)] on fm; (right plot) effect of LA on fm.
540
541

542
543 Figure 6 (upper two plots) data points of C(nLA) and 2(nLA); (lower two plots) data points of
544 C(nLA) and 2(nLA) from the twenty extra cases.

545 30
546
547 Figure 7 Simulated results for CI and SA (upper plots are for CI, and lower plots for SA) and their
548 comparisons to Eq. (13) (our model) and to spatial averaging for fD.

549

31
550
551 Figure 8 Normalized sample means of fm obtained for CI, CA, SI, and SA, and their comparisons
552 to Eq. (13) (our model)
553
554
555

32
556
557 Figure 9 Data points of C(nLA) and 2(nLA) for the QExp model
558

559

560

33
561
562
563 Figure A-1 Variations of nLA and np with respect to  for CI, CA, SI, and SA for the SExp model

564

565

566
567
568 Figure A-2 Variations of nLA and np with respect to  for CI, CA, SI, and SA for the QExp model

569

34

View publication stats