Geo-Congress 2013 ASCE 2013

Testing bias and parametric uncertainty in analyses

of a slope failure in San Francisco Bay mud
B. K. Low1, F. ASCE and J. Michael Duncan2, Hon. M. ASCE
1

Assoc Prof, School of Civil and Environmental Engineering, Nanyang Technological University,
Singapore 639798; bklow@alum.mit.edu
2
Distinguished Professor Emeritus, Department of Civil and Environmental Engineering, Virginia Tech.,
Blacksburg, VA 24061; jmd@vt.edu

ABSTRACT: An underwater slope in San Francisco Bay Mud that failed during
excavation will be analyzed, first deterministically using data from field vane shear and
laboratory triaxial tests, then probabilistically, accounting for parametric uncertainty and
correlation of the undrained shear strength and soil unit weight. In the deterministic
analysis, the factors of safety will be compared based on field vane shear test data, and
triaxial tests on 35 mm test specimens and 70 mm Shelby tube specimens. The principal
factors influencing these measures of undrained strength are shown to be sample
disturbance and testing rate. Computed factors of safety are also affected by extrapolation
of measured strengths to depths greater than the actual depths of sampling and vane shear
testing. The deterministic analysis is extended into reliability analysis to account for
parametric uncertainty and correlation among input parameters. The results from the
first-order reliability method (FORM) and Monte Carlo simulation method are compared
and discussed. The effects of different input probability distributions on the probability of
slope failure are investigated.
INTRODUCTION
The failure of a slope excavated underwater in San Francisco Bay has been described in
Duncan and Buchignani (1973), Duncan (2000, 2001), and Duncan and Wright (2005).
The slope was part of a temporary excavation and was designed with an unusually low
factor of safety to minimize construction costs. During construction a portion of the
excavated slope failed. A drawing of the slope cross section is shown in Fig. 1. The
undrained shear strength data (from in situ and lab tests) for the upper 21 m of the 30.5 m
deep excavation are presented in Fig. 2. Duncan and Buchignani calculated a short-term
factor of safety equal to 1.17 based on field vane and laboratory triaxial tests.
In this paper the slope failure is revisited, and analyzed using a reformulated
Spencers procedure of slices in an adaptable spreadsheet platform, both deterministically
using average values of input parameters, and probabilistically accounting for the
uncertainties in the values of input parameters. The reformulated Spencer procedure is
presented next, as it will be used in both the deterministic and probabilistic analyses.

937

Geo-Congress 2013 ASCE 2013

938

12

Depth m

Debris dike

6.1 m

12
30.5 m

San Francisco Bay mud

24
36
Firm soil

FIG. 1. Underwater excavated slope in San Francisco Bay mud described in Duncan
and Wright (2005)
su from UU tests on 35 mm trimmed,
kPa
0
10
20
30
40
50

su from in situ vane shear tests, kPa

10

20

30

40

5
Depth below 0.0 - m

10
15
20

10
15
20

25

25

30

30
su from UU tests on 70 mm untrimmed,
kPa
0
10
20
30
40

Depth below 0.0 - m

50

su from all data, kPa

0

50

10
15
20

Depth below 0.0 - m

Depth below 0.0 - m

10

20

30

40

50

10
15
20

25

25

30

30

FIG. 2. Undrained shear strength data from in situ vane shear and UU triaxial tests

Geo-Congress 2013 ASCE 2013

939

SPENCER METHOD REFORMULATED FOR SPREADSHEET

The derivations and procedures in this section are based on Low (2003), using the
notations in Nash (1987). The sketch at the top of Fig. 3 shows the forces acting on a
slice (slice i) that forms part of the potential sliding soil mass. The notations are: weight
Wi, base length li, base inclination angle i, total normal force Pi at the base of slice i,
mobilized shearing resistance Ti at the base of slice i, horizontal and vertical components
(Ei, Ei1, iEi, i1Ei1) of side force resultants at the left and right vertical interfaces of
slice i, where i1 and i are the tangents of the side force inclination angles (with respect
to horizontal) at the vertical interfaces. Adopting the same assumptions as Spencer
(1973), but reformulated for spreadsheet-based automatic constrained optimization
approach, one can derive the following from Mohr-Coulomb criterion and equilibrium
considerations:

Ti = ci li + (Pi ui li ) tani F

Pi cos i = Wi i E i + i 1 E i 1 Ti sin i
Ei = Ei 1 + Pi sin i Ti cos i

(Mohr-Coulomb criterion)

(1)

(Vertical equilibrium of slice i)

(2)

(Horizontal equilibrium of slice i)

(3)

Wi (i i 1 )E i 1

ci l i u i l i tan i (sin i i cos i )

Pi = F

i sin i + cos i

+ tan i (sin i i cos i )

[T cos
i

Pi sin i ] Pw = 0

(From above equations)

(4)

(Overall horizontal equilibrium) (5)

(Ti sin i + Pi cos i Wi ) L xi

+ (T cos P sin ) L M w = 0

(Overall moment equilibrium)

Lxi = 0.5( xi + xi 1 ) xc

(Horizontal lever arm of slice i) (7)

L yi = y c 0.5( y i + y i 1 )

yi

(Vertical lever arm of slice i)

(6)

(8)

where ci, i and ui are cohesion, friction angle and pore water pressure, respectively, at
the base of slice i, Pw is the water thrust in a water-filled vertical tension crack (at x0) of
depth hc, and Mw the overturning moment due to Pw. Equations 7 and 8, required for
noncircular slip surface, give the lever arms with respect to an arbitrary center. For
circular slip surfaces, Lxi = Rsini and Lyi = Rcosi, and Eq. (6) for overall moment
equilibrium reduces to: [Ti R WR sin i ] M w = 0 . The use of both i and i-1 in Eq. 4
allows for the fact that the right-most slice (slice #1) has a side that is adjacent to a
water-filled tension crack, hence 0 = 0 (i.e., the direction of water thrust is horizontal),
and for different values (either constant or varying) on the other vertical interfaces.

Geo-Congress 2013 ASCE 2013

20

940

40

60

Reformulated Spencer method

Wi

iEi

10
20

Ei-1

0.484

1.148

M
H3

li

Ei

30

i-1 Ei-1

Ti

0.00

Forces
H2
0.00

Pi
40
c0

Units: m, kN, kN/m2 (kPa), kN/m3.

Center of Rotation 2.23

slope angle

hc

48.8

30.5

0.00

10

P w Mw
0

ru

xc

0.13

yc

-20.25 56.85

Varying side-force
angle ( = tan)

1.21

xmin xmax
0.00

0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

50.50
46.32
42.65
39.41
36.56
34.06
31.85
29.91
28.20
26.70
24.79
22.90
21.03
19.18
17.35
15.54
13.74
11.96
10.20
8.46
6.73
5.02
3.33
1.66
0.00

ybot ytop b
6.10
12.88
17.49
20.84
23.39
25.36
26.92
28.17
29.18
30.01
30.97
31.84
32.62
33.31
33.93
34.47
34.95
35.35
35.70
35.99
36.22
36.39
36.51
36.58
36.60

6.10
6.10
6.10
6.10
6.10
6.10
6.10
6.10
6.10
6.10
8.28
10.44
12.57
14.69
16.78
18.85
20.91
22.94
24.95
26.94
28.91
30.86
32.79
34.71
36.60

5.72
5.72
5.72
5.72
5.72
5.72
5.72
5.72
5.72
5.72
5.72
5.72
5.72
5.72
5.72
5.72
5.72
5.72
5.72
5.72
5.72
5.72
5.72
5.72

13.7
20.6
25.4
29.0
31.7
33.9
35.6
36.9
38.0
39.1
40.2
41.2
42.1
42.9
43.6
44.2
44.8
45.2
45.6
45.9
46.2
46.3
46.4
46.5

0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

81.1
191.1
241.8
260.9
262.0
252.8
238.1
220.5
202.0
254.3
238.2
221.6
204.7
187.5
170.0
152.4
134.6
116.7
98.7
80.7
62.7
44.7
26.8
8.9

rad
1.019
0.897
0.804
0.729
0.668
0.616
0.572
0.534
0.501
0.468
0.431
0.394
0.359
0.325
0.291
0.258
0.226
0.194
0.163
0.132
0.101
0.071
0.042
0.012

x0

(xi x0 )

(xn x0 )

= sin

framed cells contain equations

xn

50.50

0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00

7.97
5.89
4.66
3.82
3.19
2.70
2.31
1.98
1.71
2.14
2.08
2.03
1.98
1.93
1.89
1.86
1.82
1.79
1.77
1.74
1.72
1.70
1.68
1.66

P
9.89
159.35
205.52
219.35
219.59
213.06
202.81
190.48
177.08
227.04
218.08
209.08
199.61
189.31
177.86
165.04
150.67
134.68
117.08
97.97
77.54
56.08
33.90
11.38

T
95.18
105.73
103.25
96.43
88.18
79.68
71.48
63.82
56.78
72.83
72.83
72.73
72.54
72.28
71.96
71.57
71.14
70.67
70.16
69.61
69.03
68.43
67.80
67.15

0
-41.5
17.0
93.3
167.6
234.3
292.3
341.9
384.0
419.3
456.6
481.4
494.6
496.9
488.8
470.9
443.8
408.2
364.8
314.6
258.4
197.6
133.3
67.0
0.0

0.000
0.124
0.227
0.308
0.369
0.413
0.443
0.463
0.475
0.482
0.483
0.478
0.467
0.450
0.426
0.398
0.365
0.328
0.287
0.243
0.197
0.149
0.100
0.050
0.000

Note: The column labels are to be interpreted in context. For the undrained analysis of the
submerged slope here, b denotes buoyant unit weight, c means cu, means u, and u is shown
as zero because it is not used in total stress analysis. Unit weight of seawater = 10 kN/m3.

FIG. 3. General template for reformulated Spencer method in spreadsheet

Geo-Congress 2013 ASCE 2013

The algebraic manipulation that results in Eq. 4 involves opening the term (Pi
uili)tani of Eq. 1, an action legitimate only if (Pi uili) > 0, or, equivalently, if the
effective normal stress i ( = Pi/li ui) at the base of a slice is non-negative. Hence, after
obtaining the critical slip surface, one needs to check that i > 0 at the base of all slices
and Ei > 0 at all the slice interfaces. Otherwise, one should consider modeling tension
cracks for slices near the upper exit end of the slip surface.
Figure 3 shows the spreadsheet set-up for deterministic stability analysis of the
underwater excavated slope in San Francisco bay. The automatic search for critical slip
surface using Microsoft Excels built-in Solver routine is described in Low (2003),
available at http://alum.mit.edu/www/bklow, including a downloadable Excel file for
hands-on to enhance understanding of the concepts and procedures.
The reformulated Spencer method (Fig. 3) enables effective stress analysis with search
for noncircular slip surfaces in heterogeneous soil characterized by shear strength
parameters c and . For the underwater slope in hand, undrained analysis is conducted,
for which the column labelled c shows the undrained shear strength, and the column
labelled shows the u values, which is 0 for saturated clay. Zeros are entered in the
column labelled u because the pore pressures are not used in total stress analysis.
Same template for Spencer, Bishop Simplified and Force Equilibrium Methods but
Different Constraints
Equations 1 to 8 are more straightforward than the original Spencer (1973) formulations.
The procedure of Fig. 3, as elaborated in Low (2003), also differs from the Spencer
method of solution, which was formulated for use in an age prior to personal computers.
The present formulation and constrained optimization procedure also allows convenient
investigations of the difference in results among Spencer, Bishops simplified, and force
equilibrium (wedge) methods. The following may be noted:
- The solution for the critical circular slip surface in Fig. 3 (and, if desired, that for the
critical noncircular slip surface) are as rigorous as the Spencer (1973) method and the
Chen-Morgenstern (1983) method (Low et al., 1998), but is operationally simpler and
conceptually more transparent.
- If cell is set to zero, and Excel Solver is invoked to change cell F, subject to the
constraint that the cell Moment be equal to 0, Solver will obtain a factor of safety
but with a net horizontal unbalanced force in the cell labelled Forces. This F
corresponds to Bishops simplified method, which assumes horizontal side-forces, and
satisfies overall moment and vertical force equilibrium but not horizontal equilibrium.
- The critical circular slip surface of Fig. 3 is for an assumed half-sine variation of the
side-force inclination parameter . If constant side-force inclination is imposed, the
factor of safety corresponding to the critical circular slip surface is also 1.15 for the
case in hand, with a constant side-force inclination parameter = 0.398.
- If Solver is invoked to change F but with fixed at some assumed value, and with
constraints on satisfying overall force equilibrium (Forces = 0) but not overall
moment equilibrium, the factor of safety obtained would correspond to the force
equilibrium (wedge) method.

941

Geo-Congress 2013 ASCE 2013

942

The following table summarizes the above procedures for conducting limit equilibrium
methods using different optimization settings.
Table 1. Implementing Spencer, Bishop Simplified and Force Equilibrium methods
on the same spreadsheet template by using different optimization settings
Method

Assumption
for or

By automatic
changing cells

Solver Constraints
regarding equilibrium

Spencer with varying

side-force inclination,
Morgenstern and Price

Varying

, F,

Forces = 0
Moments = 0

Spencer with constant

side-force inclination

Constant

, F,

Forces = 0
Moments = 0

Bishop simplified method

(Circular slip surface)

Set = 0

F,

Moments = 0

Force equilibrium or
wedge method

Example:
= (tan)/F

F,

Forces = 0

Note: and are defined in Figure 3.

DETERMINISTIC STABILITY ANALYSIS OF THE UNDERWATER SLOPE
The San Francisco Bay underwater slope was analyzed in Fig. 3 based on search for the
critical circular slip surface, using the mean value trend of all data combined (vane tests
and two types of triaxial tests, last plot of Fig. 2), for which c0 = 2.23 kPa and b = 1.21 in
the linear trend su = c0 + by. An Fs of 1.15 was obtained. Since u = 0, the same Fs will be
obtained by rotational analysis of a cylindrical block, the Ordinary Method of Slices,
Bishop Simplified Method, and Spencer circular slip surface. However, the reformulated
Spencer method affords search for the critical noncircular surface and a lower Fs, as
shown in Table 2. It is seen that while for circular slip surfaces the Fs ranges from 1.03 to
1.23, for noncircular critical slip surfaces the Fs ranges from 1.00 to 1.20. (Row 2 results
ignored, as explained later.) In other words, the Fs for the critical noncircular slip surface
is about 2.5% to 3% lower than the Fs for critical circular slip surface. Rows 2 and 3
differ in the manner of interpreting the vane shear test data. As can be seen in Figure 2,
the values of vane shear strength at 7 m depth and at 11.3 m depth are relatively very low,
and are very likely erroneous, probably as a result of having been performed in mud that
was disturbed in the process of drilling the borehole or inserting the vane. Row 2 of Table
2 reflects calculations that include these two tests. Row 3 of Table 2 reflects calculations
in which these two values were excluded. Paradoxically, eliminating the low strengths
had the effect of reducing the computed factor of safety. This illogical result is a
consequence of extrapolating the measured strengths from the upper 21 m of Bay mud,
where the tests were performed, to 31 m, which is the bottom of the Bay mud deposit.
This is discussed more completely in the following section.

Geo-Congress 2013 ASCE 2013

943

Table 2. Deterministic analysis based on mean values of su, using Spencer method
for circular and noncircular slip surfaces.
c0

Tests

(kPa)

All data combined

In situ vane shear including two bad low-values
In situ vane shear excluding two bad tests
35 mm trimmed triaxial test specimens
70 mm untrimmed triaxial test specimens

2.23
2.71
9.11
1.10
0.05

Slope of
trend
line, b

Fs, critical
circular
surface

Fs, critical
noncircular
surface

1.21
1.35
1.04
1.30
1.17

1.15
1.29
1.23
1.19
1.03

1.12
1.26
1.20
1.16
1.00

Figure 4 shows the critical noncircular slip surface for the all-data case, obtained
using Microsoft Excels built-in constrained optimization Solver routine, by varying 24
degrees of freedom which eventually deforms the critical circular surface of Fig. 3 into
the critical noncircular surface of Fig. 4. The su profile is the mean trend of all data, for
which su = 2.23 + 1.21y, in kPa, where y is the depth in meter below 0.0.
r0
0.95
0

su - kPa
10

20

30

40

50

60

10

10

15

15

y - meter

20

30
35
40

20

30

40

50

25

10

Critical noncircular slip

surf ace: Fs = 1.12,
based on the mean su
trend of all data.

20
25
30
35
40

Stiff layer

FIG. 4. The reformulated Spencer method obtains the critical noncircular slip
surface as shown and a factor of safety of 1.12, compared with a Fs of 1.15 based on
critical circular slip surface.
BIASES IN STRENGTH MEASUREMENTS AND INTERPRETATIONS
The results of both the deterministic and the probabilistic analyses are affected by biases
in the strength measurements and interpretations. The sources of these biases include
disturbance resulting from sampling and handling the triaxial test specimens and from
inserting the vanes; rates of shearing in the triaxial and vane shear tests that are much
more rapid than in the field; and extrapolating the measured strengths to the full depth of
the deposit. These three sources of bias are discussed below.
Disturbance Effects. It is well understood that inserting a tube sampler below the
bottom of a borehole disturbs the soil to some extent, even if a thin-walled sampler is
used. Similarly, inserting a vane below the bottom of the borehole also causes some
amount of disturbance, and in both cases the effect of disturbance is to reduce the

Geo-Congress 2013 ASCE 2013

undrained strength of a normally consolidated clay like San Francisco Bay mud. In the
site investigation for the LASH terminal, an attempt was made to minimize the amount of
disturbance in some of the triaxial test specimens by trimming away the outer 18 mm of
the 70-mm Shelby tube samples, because it was thought that the outer part of the samples
would be more disturbed than the inner part which is further from the wall of the
sampling tube. Other test specimens were prepared by extruding the entire 70-mm
samples. It can be noted in Figure 2 that the strengths measured using the 35-mm
trimmed test specimens were consistently higher than those measured using 70-mm
(untrimmed) test specimens. The effect of disturbance reducing undrained strength is
reflected in the fact that the minimum factors of safety calculated using the data from the
tests on 35-mm specimens are about 15 percent higher than those calculated using data
from the 70-mm specimens. While it can be seen that trimming away the outer portion of
the tube samples served to reduce the strength loss due to disturbance, it is likely that it
did not completely eliminate the effect of disturbance.
Rates of testing. Undrained strengths of saturated clays like San Francisco Bay mud
depend to an important extent on the length of time required to reach failure. Duncan and
Buchignani (1973) showed that the undrained strengths measured in tests where the time
to failure was a week or so was about 30 percent lower than in tests where the time to
failure was 10 to 20 minutes. The triaxial tests performed on both the 35-mm and the
70-mm test specimens were performed at normal rates of loading, where failure was
reached in 10 to 20 minutes. The vane shear tests were also performed at normal rates of
rotation of the vane (one degree of rotation per 10 seconds), and failure was reached in
one to two minutes. Thus the Bay mud was brought to failure in both the triaxial tests and
the vane shear tests in much less time than the excavated slope at the LASH terminal was
expected to remain stable. This shorter time to failure would increase the measured
strength as compared to that mobilized in the field, where failure occurred after a period
of days or weeks.
Extrapolation of Measured Strengths to Greater Depths. Due to the difficulties in
sampling and performing vane shear tests at greater depths with the equipment available
in 1970, tests were not performed below a depth of 21 m beneath the mudline. The fact
that su/p is constant for normally consolidated clays like San Francisco Bay mud had been
well established, and was supported by the increase in strengths down to 21 m. It seemed
logical to expect that the strength should continue to increase linearly below 21 m, and it
still does. However, extrapolating the strengths measured above a depth of 21 m to the
bottom of the profile resulted in a biased strength profile because of a secondary
influence of disturbance. Both sampling and vane shear testing became more difficult,
and probably resulted in greater disturbance effects as the depth increased. This would
result in too small a rate of increase of strength with depth, which would affect all of the
measured strengths. Ironically, the factor of safety calculated for the linear variation of
strength with depth with two low values of vane shear strength eliminated resulted in a
lower factor of safety than when the erroneous low strengths were included. The reason
for this paradoxical result is that the erroneous low strengths were at relatively shallow
depths. Eliminating these low strength values resulted in a higher average strength within
the upper 15 m of the Bay mud, but it also reduced the apparent rate of strength increase
with depth. The average strength over the full 31 m thickness of the Bay mud turned out
to be about the same with and without the two low value strengths, as can be seen in

944

Geo-Congress 2013 ASCE 2013

945

Figure 5. However, because the portion of the slip surface in the lower 16 m of the Bay
mud is significantly longer than the portion in the upper 15 m (Figs. 3 and 4), the
extrapolated su profile with the two erroneous low strengths omitted results in a lower
factor of safety than if the two erroneous low strengths were not removed. It is interesting
that using a nonlinear trend line for the variation of strength with depth, as shown at the
right in Figure 5, does not lead to the paradox that eliminating low strengths results in
lower factors of safety. However, this nonlinear variation of strength with depth is not
consistent with the expected behavior of saturated clays like San Francisco Bay mud, and
hence will not be used for the case in hand.
Importance of Judgment. The preceding discussion shows that none of the means of
measuring the undrained strength is perfect. All involve bias of one type or another:
lower strength due to disturbance, higher strength due to faster rate of testing, and
complex effects due to extrapolating strengths beyond the depths where they were
measured. As a result, in this situation as in many others, geotechnical engineers face the
necessity of working with imperfect data. Probabilistic analyses are helpful for
understanding the effects of random variations in measured properties, but even
probabilistic analyses are affected by biases in the underlying data. To cope with the
effects of bias requires an understanding of the behavior of the soil, and the application of
engineering judgment. Engineering judgment indicates, for example, that a factor of
safety of 1.17 is too low for a project like the LASH Terminal excavated slope.
su from in situ vane shear tests, kPa

su from in situ vane shear tests, kPa

0

15

30

45

60

30

45

Depth below 0.0 - m

Depth below 0.0 - m

5
10
15
20
25
30

10
15
20
25

35
40

15

Inflexible linear trend plus extrapolation

lead to misleading su trend and higher F s .

e data
30

Ignore two bad points

FIG. 5. Excavation is from depth 6.1 m to 36.6 m (a) Linear regression in the top 2/3
of the 30.5 m excavation, and extrapolation into the lower 1/3 of the excavation. (b)
Second- degree polynomial model for the top 2/3 of the excavation.
Probabilistic analyses are conducted in the following sections on three sets of data for
which the linear su-profile assumption seems to be good approximations (at least for the
depth where su data are available), namely (i) field vane test data with two bad points
eliminated, (ii) UU triaxial test data on 35 mm specimens trimmed from 70 mm Shelby
tube samples, (iii) UU triaxial test data on untrimmed 70 mm specimens. The mean-value
trend and the 1 standard deviation lines for the three types of data are shown in Fig. 6.

Geo-Congress 2013 ASCE 2013

946

The undrained shear strength profile is modeled as a random variable by

su = (c0 + by )

(9)

where y = depth in meter (> 6.1 m), is the ratio of measured su values to trendline
values. The values of c0 and b are given in Fig. 6. The mean of is equal to 1. The
standard deviations of are the values in Fig. 6.
Another quantity treated as a random variable is the buoyant unit weight of the Bay
mud, with a mean value of 5.72 kN/m3 and a standard deviation of 0.52 kN/m3.
s u f rom UU tests on 35 mm
trimmed, kPa

s u f rom in situ vane shear

tests, kPa
15

30

45

c 0 = 9.11
b = 1.04
= 0.18

30

45

10
15

c 0 = 1.10
b = 1.30
= 0.21

10
15

10

45

c 0 = 0.05
b = 1.17
= 0.21

15
20

25

25

25

30

30

30

Two bad
points
removed

30

20

20

15

Depth below 0.0 - m

Depth below 0.0 - m

15

Depth below 0.0 - m

s u f rom UU tests on 70mm

untrimmed, kPa

FIG. 6. The average su profile, described by su = c0 + by, and the 1 Std. Dev. lines.
EFFICIENT APPROACH FOR FIRST-ORDER RELIABILITY METHOD
The Hasofer-Lind index (for correlated normal random variables) and the first-order
reliability method (FORM, for correlated nonnormals) can be better understood by
adopting an intuitive perspective and a spreadsheet-based reliability approach as
described in Low and Tang (2004, 2007), Low (2008) and Low et al. (2007). The
procedures can be applied to stand-alone numerical (e.g. finite element) packages via the
response surface method. Hence the applicability of the reliability approach is not
confined to models which can be formulated in the spreadsheet environment.
The established matrix formulation of the Hasofer-Lind index is:

= min
xF

(x )T C 1 (x )

(10a)

or, equivalently:

x i 1 xi i
= min i
R
xF
i

(10b)

where x is a vector representing the set of random variables xi, the vector of mean
values i, C the covariance matrix, R the correlation matrix, i the standard deviations,

Geo-Congress 2013 ASCE 2013

947

and F the failure domain. Low and Tang (1997) used Eq. (10b) instead of Eq. (10a),
because the correlation matrix R is easier to set up, and conveys the correlation structure
more explicitly than the covariance matrix C.
The point denoted by the xi values, which minimize Eq. (10) and satisfies x F is
the design point. This is the point of tangency of an expanding dispersion ellipsoid with
the limit state surface (LSS), which separates safe combinations of parametric values
from unsafe combinations. The quadratic form in Eq. (10) appears also in the negative
exponent of the established probability density function of the multivariate normal
distribution. As a multivariate normal dispersion ellipsoid expands from the mean-value
point, its expanding surfaces are contours of decreasing probability values. Hence, to
obtain by Eq. (10) means maximizing the value of the multivariate normal probability
density function, and is graphically equivalent to finding the smallest ellipsoid tangent to
the LSS at the most probable failure point (the design point).
For correlated nonnormals, the ellipsoid perspective still applies in the original
coordinate system, except that the nonnormal distributions are replaced by an equivalent
normal ellipsoid, centered not at the original mean values of the nonnormal distributions,
but at the equivalent normal mean-value point. Details are in Low and Tang (2004).
An alternative FORM computational procedure is given in Low & Tang (2007),
which uses the following equation for the reliability index :

= min n T R 1n
xF

(11)

The reliability index and the numerical values of the dimensionless equivalent standard
normal vector n at the design point (denoted as n*) are obtained via
spreadsheet-automated constrained optimization. The original random variables xi (on
which the performance functions are formulated) are computed automatically from the
probabilistic connections between ni and xi.
RELIABILITY ANALYSIS OF THE SLOPE IN SAN FRANCISCO BAY MUD
The computation and concepts will be illustrated first for uncorrelated normal random
variables, then for correlated normal random variables, and finally for correlated
lognormals.
Simple illustration of FORM for uncorrelated normal random variables and b
The FORM reliability index is obtained easily using the Low and Tang (2007)
spreadsheet procedure, with search for the reliability-based critical slip circle. The
probability of failure is then computed from Pf = 1 () = 1 NormSDist().
Comparisons are made with three Monte Carlo simulations each with 3000 trials and
different initial seed, using the commercial software @RISK (http://www.palisade.com),
based on the reliability-based critical slip circle. The three Monte Carlo Pf values are
shown below the tables for comparison with the FORM Pf values. The Pf values from
reliability index are practically identical to the Pf from Monte Carlo simulations. The
computation time for each FORM is only a second or two. The computation time for
Monte Carlo simulation is about 500 times higher.

Geo-Congress 2013 ASCE 2013

948

Table 3a. In situ vane shear excluding two bad tests

Correlation matrix
Distribution
x
n

Normal
Normal

b 5.72

0.18

0.836

-0.913

0.52

5.914

0.373

0.99

Pf
16.2%

Compare: Pf from three Monte Carlo simulations: 16.1%, 15.9%, 16.1%.

Table 3b. 35 mm diameter UU triaxial trimmed specimens

Correlation matrix
Distribution
x
n

Normal
Normal

b 5.72

0.21

0.859

-0.669

0.52

5.847

0.243

0.71

Pf
23.8%

Compare: Pf from three Monte Carlo simulations: 23.6%, 24.3%, 23.4%.

Table 3c. 70 mm diameter UU triaxial untrimmed specimens

Correlation matrix
Distribution
x
n

Normal
Normal

b 5.72

0.21

0.969

-0.149

0.52

5.752

0.062

0.16

Pf
43.6%

Compare: Pf from three Monte Carlo simulations: 44.0%, 43.5%, 43.2%.

FORM analysis for correlated normal random variables and b

The undrained shear strength su in Eq. (9) increases with increase in the parameter .
Physical considerations then suggest that the parameter should be positively correlated
with the buoyant unit weight b, because high su values are likely to go with high unit
weights, and vice versa. In this section reliability analysis is performed assuming a
positive correlation coefficient of 0.5 between and b. The results are shown below. It
can be seen that when a strength parameter () is positively correlated with a load
parameter (b), the probability of failure is lower than the uncorrelated case of Table 3.
Table 4a. In situ vane shear excluding two bad tests
Correlation matrix
Distribution
x
n

Normal
Normal

b 5.72

0.18

0.799

-1.118

0.5

0.52

5.653

-0.129

0.5

1.22

Pf
11.1%

Compare: Pf from three Monte Carlo simulations: 11.6%, 11.5%, 11.1%.

Table 4b. 35 mm diameter UU triaxial trimmed specimens

Correlation matrix
Distribution
x
n

Normal
Normal

b 5.72

0.21

0.831

-0.807

0.5

0.52

5.650

-0.134

0.5

0.86

Pf
19.4%

Compare: Pf from three Monte Carlo simulations: 18.6%, 19.8%, 19.4%

Geo-Congress 2013 ASCE 2013

949

Table 4c. 70 mm diameter UU triaxial untrimmed specimens

Correlation matrix
Distribution
x
n

Normal
Normal

b 5.72

0.21

0.961

-0.183

0.5

0.52

5.710

-0.019

0.5

0.20

Pf
42.0%

Compare: Pf from three Monte Carlo simulations: 42.5%, 41.8%, 41.6%.

FORM analysis for correlated lognormal random variables and b

Input random variables are often modeled by the lognormal distribution, which provides
some mathematical convenience and also avoids the negative domain. For the case in
hand where the coefficient of variation is 0.21, the failure probabilities for correlated
lognormals (Table 5) do not differ much from those for correlated normals (Table 4).
Table 5a. In situ vane shear excluding two bad tests
Correlation matrix
Distribution
x
n

Lognormal
Lognormal

b 5.72

0.18

0.806

-1.120

0.5

0.52

5.703

0.012

0.5

1.30

Pf
9.7%

Compare: Pf from three Monte Carlo simulations: 9.80%, 9.63%, 10.1%.

Table 5b. 35 mm diameter UU triaxial trimmed specimens

Correlation matrix
Distribution
x
n

Lognormal
Lognormal

b 5.72

0.21

0.833

-0.778

0.5

0.52

5.664

-0.063

0.5

0.86

Pf
19.4%

Compare: Pf from three Monte Carlo simulations: 19.4%, 19.8%, 19.2%

Table 5c. 70 mm diameter UU triaxial untrimmed specimens

Correlation matrix
Distribution
x
n

Lognormal
Lognormal

b 5.72

0.21

0.959

-0.100

0.5

0.52

5.692

-0.008

0.5

0.11

Pf
45.6%

Compare: Pf from three Monte Carlo simulations: 45.2%, 46.3%, 45.9%.

It seems logical to assume that the strain rate effect is more dominant than the
disturbance effect in the field vane tests and in the UU triaxial tests on 35 mm trimmed
specimens, thereby resulting in overestimated undrained shear strengths and hence
underestimated probabilities of failure. In any case, a computed (and underestimated) Pf
of about 10% (Table 5a, lowest of Tables 3, 4 and 5) would have made the slope design
unacceptably risky, since reliability-based design typically aims at a of at least 2.5 (i.e.,
Pf < 1%).
SUMMARY AND CONCLUSIONS
An underwater slope excavated in San Francisco Bay mud has been analyzed first
deterministically using a reformulated Spencer method, then probabilistically using the

Geo-Congress 2013 ASCE 2013

first-order reliability method (FORM). Factors of safety and reliability indices are
computed separately for three types of measured undrained shear strength data, namely
field vane tests, UU triaxial tests on trimmed 35 mm specimens, and UU triaxial tests on
untrimmed 70 mm specimens. The principal factors affecting the measured strength
values are disturbance and rate of loading effects, together with subtle errors caused by
extrapolating measured strength data into untested lower depth of the Bay mud.
On the basis of this study, the following conclusions are justified:
(i) The results of both the deterministic and the probabilistic analyses are affected by
biases in the strength measurements and interpretations, and by extrapolating the
measured strengths to the full depth of the deposit.
(ii) To cope with the effects of bias requires an understanding of the behavior of the soil,
and the application of engineering judgment in both the deterministic and the
probabilistic analysis.
(iii)For the slope in hand, the computed factors of safety are 1.20, 1.16 and 1.00, based on
field vane, trimmed 35 mm diameter and untrimmed 70 mm diameter specimens in
UU triaxial tests. The FORM analyses produce probabilities of failure between 10%
and 46%. An unusually low factor of safety = 1.17 was deemed acceptable for design
of this slope because good quality tests had been performed and because the
excavated slope only had to be stable temporarily. Conventional design would have
required a higher factor of safety (about 1.3) at the end of construction.
(iv) Given the fact that the slope failed, the right factor of safety is 1.0. It is ironic that
the strength measurements that result in a factor of safety closest to 1.0 are those that
were considered to be of lowest quality the tests performed on the 70 mm samples,
where the most disturbed outer portion of the samples was not trimmed away.
However, for these samples, the effect of the greater degree of disturbance reducing
the undrained strengths was compensated very closely by the rate of loading
increasing the undrained strengths. This shows clearly the significant effect of rate of
loading, which is commonly ignored.
(v) That judgment is required in estimating the statistical inputs and in interpreting the
outputs of a probabilistic analysis should not deter one from adopting the probabilistic
approach (which is more rational and has richer output information), but should serve
as an incentive to find ways to improve the approach, including more thought given
to estimation of statistical inputs and their probabilistic modeling.

REFERENCES
Baecher, G.B. and Christian, J.T. (2003). Reliability and Statistics in Geotechnical
Engineering. Wiley: Chichester, UK; Hoboken, NJ, 618 p.
Baecher, G.B. and Christian, J.T. (2008). Spatial variability and geotechnical reliability,
Chapter 2 (pages 76-133) of Reliability-Based Design in Geotechnical EngineeringComputations and Applications, Taylor and Francis, ed. K.K.Phoon.
Christian, J.T., Ladd, C.C. and Baecher, G.B. (1994). Reliability applied to slope
stability analysis. J. Geotech. Eng. ASCE, 120(12): 2180-2207.
Duncan, J.M. (2000). Factors of safety and reliability in geotechnical engineering. J.
Geotechnical and Geoenvironmental Engrg. 126(4): 307-316.

950

Geo-Congress 2013 ASCE 2013

Duncan, J.M. (2001). Closure to Discussions on Factors of safety and reliability in

geotechnical engineering. J. Geotechnical and Geoenvironmental Engrg. 127(8):
717721.
Duncan, J.M. and Buchignani, A.L. (1973). Failure of underwater slope in San
Francisco Bay. J. Soil Mechanics and Foundation Division, ASCE, 99(9): 687-703.
Duncan, J.M. and Wright, S.G. (2005). Soil strength and slope stability. Wiley & Sons,
297 p.
Low, B.K. (2003). Practical probabilistic slope stability analysis. Proceedings, Soil and
Rock America 2003, Cambridge, Massachusetts, USA, June 22-26, 2003, Verlag
Glckauf GmbH Essen, Vol. 2, 2777-2784.
Low, B.K. (2008). Practical reliability approach using spreadsheet, Chapter 3 (pages
134-168) of Reliability-Based Design in Geotechnical Engineering-Computations and
Applications, Taylor and Francis, ed. K.K.Phoon.
Low, B.K. and Wilson H. Tang (1997). "Reliability analysis of reinforced embankments
on soft ground." Canadian Geotechnical Journal, 34(5): 672-685.
Low, B.K., and Wilson H. Tang (2004). Reliability analysis using object-oriented
constrained optimization. Structural Safety, Elsevier, Amsterdam, 26(1), 69-89.
Low, B.K. and Wilson H. Tang (2007). Efficient spreadsheet algorithm for first-order
reliability method. J. Engineering Mechanics, ASCE, 133(12): 1378-1387.
Low B.K., Gilbert, R.B. and Wright, S.G. (1998). "Slope reliability analysis using
generalized method of slices." J. Geotech. and Geoenviron. Engrg. 124(4): 350-362.
Low, B.K., Lacasse, S. and Nadim, F. (2007). Slope reliability analysis accounting for
spatial variation. Georisk: Assessment and Management of Risk for Engineered
Systems and Geohazards, Taylor & Francis, London, 1(4): 177-189.
Morgenstern, N. R., and Price, V. E. (1965). The analysis of the stability of general slip
surfaces, Geotechnique 15(1), 79-93.
Nash, D., A comparative review of limit equilibrium methods of stability analysis. In M.
G. Anderson and K. S. Richards (Eds), Slope Stability, 1987, 11-75, Wiley: New York.
Spencer, E. (1973). Thrust line criterion in embankment stability analysis.
Geotechnique, 1973, 23: 85-100.
Vanmarcke, E.H. (1977). Reliability of earth slopes. J. Geotech. Eng. ASCE, 103(11):
1247-1266.
Vanmarcke, E.H. (1980). Probabilistic stability analysis of earth slopes. Eng. Geol.: 16,
29-50.

951