Você está na página 1de 15

Computers and Geotechnics 65 (2015) 30–44

Contents lists available at ScienceDirect

Computers and Geotechnics


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

Research Paper

Reliability-based design and its complementary role to Eurocode 7


design approach
Bak Kong Low a,⇑, Kok-Kwang Phoon b
a
School of CEE, Nanyang Technological University, Singapore
b
Department of CEE, National University of Singapore, Singapore

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

Article history: Reliability-based design (RBD) can play a useful complementary role to overcome some limitations in the
Received 29 July 2014 Eurocode 7 (EC7) design approach, for example in situations with parameters not covered in EC7, differ-
Received in revised form 19 October 2014 ent parametric sensitivities across different problems, cross-correlated or spatially correlated parameters,
Accepted 23 November 2014
design aiming at a target reliability or failure probability, or when uncertainty in unit weight of soil is
modeled. The complementary role played by RBD under these circumstances is illustrated and discussed
for a shallow foundation, a reinforced rock slope, a Norwegian clay slope with spatial variability, a later-
Keywords:
ally loaded pile requiring implicit numerical analysis, and an anchored sheet pile wall. A pragmatic RBD
Reliability-based design
Reliability index
approach involving first-order reliability method (FORM) only and a more rigorous RBD approach involv-
Eurocode 7 ing both first-order and second-order reliability method (SORM) are offered. Both approaches are imple-
Slopes mentable using either spreadsheet-based FORM and SORM procedures, or using various commercially
Earth retaining walls available FORM/SORM packages.
Laterally loaded pile Ó 2014 Elsevier Ltd. All rights reserved.

1. Introduction Hasofer–Lind index [19] for correlated normal random variables.


These reliability methods are described in Ditlevsen [15], Ang
The design approach based on overall factor of safety has long and Tang [1], Madsen et al. [36], Haldar and Mahadevan [18], Mel-
been used by geotechnical engineers. More recent alternatives chers [38], Baecher and Christian [2], for example. In addition, Low
are the characteristic values and partial factors used in the limit and Tang [22,26,30] presented spreadsheet-based practical proce-
state design approach in Eurocode 7 (EC7), and the load and resis- dures for FORM reliability-based analysis and design, with the
tance factor design (LRFD) approach in North America. Yet another 2007 approach being an equally efficient alternative to the 2004
approach can play at least a useful complementary role to EC7 and approach, and the 2004 approach much more versatile and effi-
LRFD, namely the design based on a target reliability index that cient than the 1997 prototype.
explicitly reflects the uncertainty of the parameters and their cor- This study illustrates geotechnical reliability-based design/anal-
relation structure. Although uncertainties have been considered in ysis for a spread footing, a reinforced rock slope, a soil slope, a
EC7 and LRFD in an approximate way, correlations (both spatial laterally loaded pile, and an anchored sheet-pile wall. Focus is on
correlation between the same parameter at different sampling practical procedures in the Excel spreadsheet platform for FORM,
points and cross-correlation between different parameters at the and discussions of interesting insights which suggest that FORM
same sampling point) are not considered. Extensive soil databases reliability-based design can play at least a complementary role to
have been compiled to demonstrate the presence of cross-correla- the Eurocode 7 design approach. In FORM the design point is a
tions between soil parameters fairly conclusively [8,9,11,12,10]. point on the boundary (the limit state surface) which separates
The latest database presented by Ching and Phoon [11] contains safe combinations of parametric values (e.g. the mean-value point)
7490 sets of data, each set consisting of 10 clay parameters mea- from the unsafe combinations of parametric values (e.g. occurrence
sured around the same sampling point. of low strength in combination with a large loading). The design
Among the various versions of reliability indices, that based on point is the most-probable combination of parametric values at
the first-order reliability method (FORM) for correlated nonnor- failure. The similarities and differences between the ratios of mean
mals is most consistent. A special case of FORM is the earlier values to design-point values on the one hand, and the partial
factors and the characteristic values of limit state design in
Eurocode 7 on the other hand, will be discussed.
⇑ Corresponding author.

http://dx.doi.org/10.1016/j.compgeo.2014.11.011
0266-352X/Ó 2014 Elsevier Ltd. All rights reserved.
B.K. Low, K.-K. Phoon / Computers and Geotechnics 65 (2015) 30–44 31

A brief summary is given next for the two relatively intuitive (LSS), which separates safe combinations of parametric values from
and transparent FORM computational procedures of Low and Tang unsafe combinations (Fig. 1). The one-standard-deviation (1-r) dis-
[26,30], and the associated dispersion ellipsoid perspective in the persion ellipse and the b-ellipse in Fig. 1 are tilted by virtue of cohe-
original space of random variables, aiming at overcoming the lan- sion c and friction angle / being negatively correlated. The
guage and conceptual barriers (aptly noted by Whitman [46]) sur- quadratic form in Eq. (1) appears also in the negative exponent of
rounding reliability analysis. The physical meaning of the design the established probability density function of the multivariate nor-
point obtained in FORM is explained, and it will be shown that it mal distribution. As a multivariate normal dispersion ellipsoid
is related to the characteristic values and partial factors of EC7. expands from the mean-value point, its expanding surfaces are con-
(Guidance on EC7 can be found in EN1997-1 [16], Simpson and tours of decreasing probability values. Hence, to obtain b by Eq. (1)
Driscoll [42], and Frank et al. [17], for example.) means maximizing the value of the multivariate normal probability
density function (at the most probable failure combination of para-
metric values), and is graphically equivalent to finding the smallest
2. Practical FORM procedures and intuitive dispersion ellipsoid
ellipsoid tangent to the LSS at the most probable failure point (the
perspective
design point). This intuitive and visual understanding of the design
point is consistent with the more mathematical approach in Shin-
The merits of reliability-based design (RBD) and how it can
ozuka [40], in which all variables were standardized and the limit
complement EC7 design approach are the focus of Section 4 and
state equation was written in terms of standardized variables.
later sections. A prerequisite for RBD to play a complementary role
In FORM (for random variables obeying non-normal distribu-
to the EC7 design approach is that the RBD approach must be rel-
tions), one can rewrite Eq. (1b) as follows [26], and regard the com-
atively transparent and its meaning readily appreciated by engi-
putation of b as that of finding the smallest equivalent
neers, and a balance must thus be sought between rigorous RBD
hyperellipsoid (centered at the equivalent normal mean-value
and pragmatic RBD. Hence the concepts and procedures of RBD
point lN and with equivalent normal standard deviations rN) that
are explained in this section and the next.
is tangent to the LSS:
The matrix formulation [45,15] of the Hasofer–Lind [19] index b
is: sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 T  ffi
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi xi  lNi 1 xi  li
N
b ¼ min R ð2Þ
b ¼ min ðx  lÞT C1 ðx  lÞ ð1aÞ x2F N
ri riN

x2F
where lNi and rNi can be calculated by the Rackwitz–Fiessler [39]
where x is a vector representing the set of random variables xi, l the
transformation. Hence, for correlated nonnormals, the ellipsoid per-
vector of mean values li, C the covariance matrix, and F the failure
spective still applies in the original coordinate system, except that
domain. The notations ‘‘T’’ and ‘‘1’’ denote matrix transpose and
the nonnormal distributions are replaced by an equivalent normal
matrix inverse, respectively. The symbol ‘‘2’’ means ‘‘belongs to’’.
ellipsoid, centered not at the original mean values of the nonnormal
The following alternative formulation, which is mathematically
distributions, but at the equivalent normal mean lN.
equivalent to Eq. 1(a), was used in Low and Tang [23], because the
Eq. (2) and the Rackwitz–Fiessler equations (for lNi and rNi )
correlation matrix R is easier to set up, and conveys the correlation
were used in the spreadsheet-automated constrained optimization
structure more explicitly than the covariance matrix C:
FORM computational approach in Low and Tang [26]. An alterna-
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi tive to the 2004 FORM procedure is given in Low and Tang [30],
 T  
xi  l i xi  li
b ¼ min R1 ð1bÞ which uses the following equation for the reliability index b:
x2F ri ri pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
b ¼ min nT R1 n ð3Þ
where R is the correlation matrix, ri the standard deviations, and x2F
other symbols as defined for Eq. (1a). It is clear from Eq. (1b) that
standard deviations and correlations are explicitly considered in where n stands for the dimensionless vector ðxi  lNi Þ=rNi of Eq. (2).
For each value of ni tried by Excel Solver, a short and simple Excel
the form of ri and R, respectively. The point denoted by the xi val-
ues, which minimizes Eq. (1) and satisfies x 2 F (i.e., x belongs to VBA code automates the computation of xi from ni for use in the
constraint g(x) = 0, via xi ¼ F 1 ½Uðni Þ, where U is the standard nor-
the failure domain), is the design point. This is the point of tangency
of an expanding dispersion ellipsoid with the limit state surface mal distribution and F is the original nonnormal distribution,
thereby obviating the need to compute lNi and rNi .
The computational approaches of Eqs. (1b), (2) and (3) and the
β = R/r associated ellipsoidal perspective are complementary to the classi-
cal u-space computational approach, and may help reduce the con-
Limit state surface: boundary between ceptual and language barriers of FORM. Note that the variables in
safe and unsafe domains u-space are uncorrelated standard normals while the variables in

SAFE σφ One-sigma
n-space are correlated standard normals as explained below. It is
also clear that Eq. (1) is a special case (normal distributions) of
Cohesion, c

dispersion ellipse
Eq. (2).
β-ellipse The two spreadsheet-based computational approaches of
μc FORM, Eqs. (2) and (3), are compared in Fig. 2. Either method can
r σc be used as an alternative to the classical u-space FORM procedure.
UNSAFE R
A third alternative (illustrated in [35] is also shown in Fig. 2, for
which the Microsoft Excel’s built-in constrained optimization rou-
Design Mean-value
point
tine (Solver) is invoked to vary the u vector automatically so that b
point
μφ and the design point are obtained. This requires only adding one u
column to the 2007 procedure, and expressing the unrotated n
Friction angle, φ
vector in terms of the u, where u is the uncorrelated standard
Fig. 1. Illustration of the reliability index b in the plane when c and / are negatively equivalent normal vector in the rotated space of the classical
correlated. mathematical approach of FORM. The vectors n and u can be
32 B.K. Low, K.-K. Phoon / Computers and Geotechnics 65 (2015) 30–44

Fig. 2. Comparison of the two FORM computational approaches of Low and Tang [26,30], and the additional u-to-n-to-x approach. All three procedures are open-source and
use the optimization routine Solver resident in the Microsoft Excel spreadsheet.

obtained from one another, n = Lu and u = L1n, as follows (e.g., founded at a depth of 1.8 m in a silty sand with friction angle /
[33]): = 25°, cohesion c = 15 kN/m2, and unit weight c = 21 kN/m3.
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi With respect to bearing capacity failure, the performance func-
T
b ¼ min nT R1 n ¼ min nT ðLUÞ1 n ¼ min ðL1 nÞ ðL1 nÞ tion (PerFunc) is:
x2F x2F x2F

ð4aÞ PerFunc ¼ qu  q ð6aÞ

pffiffiffiffiffiffiffiffiffi B0
i:e: b ¼ min uT u; where u ¼ L1 n; and n ¼ Lu ð4bÞ where qu ¼ cN c sc dc ic þ po Nq sq dq iq þ cNc sc dc ic ;
x2F 2
q ¼ Q v =B0 ð6b; cÞ
in which L is the lower Cholesky triangular matrix of R. When the
random variables are uncorrelated, u = n by Eq. (4), because then in which qu is the ultimate bearing capacity, q the applied bearing
L1 = L = I (the identity matrix). pressure, c the cohesion of soil, po the effective overburden pressure
The probability of failure can be estimated approximately as at foundation level, B0 the effective width of foundation (B0 = B  2e,
follows: where e = load eccentricity), c the unit weight of soil below the base
Pf  1  UðbÞ ¼ UðbÞ ð5Þ of foundation, and Nc, Nq, and Nc are bearing capacity factors, which
are functions of the friction angle (/) of soil:
where U is the cumulative distribution (CDF) of the standard nor-  
/
mal random variable, computed easily using the formula ‘‘ = Nor- Nc ¼ ðNq  1Þ cotð/Þ; Nq ¼ ep tan / tan2 45 þ ;
2
mSDist(b)’’ in Microsoft Excel.  
Some files illustrating the [26,30] open-source approaches are Nc ¼ 2 Nq þ 1 tan / ð7a; b; cÞ
available at http://alum.mit.edu/www/bklow. Step-by-step guid- Several expressions for Nc exist. The above Nc is attributed to
ance which enables hands-on appreciations of the [26] spreadsheet Vesic in Bowles [3]. The nine factors sj, dj and ij in Eq. (6b) account
procedure is available in the Low [32] chapter, and will not be for the shape and depth effects of foundation and the inclination
repeated here. Instead, the following sections illustrate the Low effect of the applied load. The formulas for these factors are based
and Tang [30] spreadsheet procedure for FORM, and discuss on Tables 4.5a and 4.5b of Bowles [3].
insights, advantages and subtleties associated with FORM reliabil- To illustrate reliability-based design, Tomlinson’s deterministic
ity-based design in various geotechnical engineering examples, example is extended probabilistically here (Fig. 3). The width B of
and the complementary role which FORM reliability-based design the foundation is to be determined based on a reliability index
can play to design based on Eurocode 7. (Although this study uses b = 3.0 against bearing capacity failure. The parameters c, /, Qh,
the [30] procedure, the same results will be obtained if the [26] and Qv are assumed to be lognormal random variables with mean
procedure is used.) values equal to the values in Tomlinson’s deterministic example,
and with coefficient of variation equal to 0.20, 0.1, 0.15, 0.10,
3. Practical spreadsheet procedure illustrated for the reliability- respectively. The mean and standard deviation of these four vari-
based design of a shallow foundation ables are shown in Fig. 3. The random variables are partially corre-
lated, with correlation matrix as shown in the figure. The
Tomlinson [44] Example 2.2 determines the factor of safety foundation width required to achieve a reliability index b of 3.0
against bearing capacity failure of a retaining wall that carries a is B = 4.51 m, using the Low and Tang [30] FORM computational
horizontal load (Qh) of 300 kN/m applied at a point 2.5 m above approach on the Microsoft Excel spreadsheet platform. Although
the base and a centrally applied vertical load (Qv) of 1100 kN/m, closed-form bearing capacity equations are used in the example
as shown in Fig. 3. The base (5 m  25 m) of the retaining wall is of Fig. 3, in other cases/problems the practical FORM procedure
B.K. Low, K.-K. Phoon / Computers and Geotechnics 65 (2015) 30–44 33

Fig. 3. Determining foundation width B, for a reliability index b of 3.0, using the Low and Tang [30] approach via Microsoft Excel Solver optimization tool.

below can be coupled with stand-alone implicit and/or iterative case in hand, g(x) displays ‘‘21.84’’ if trial B = 3, and ‘‘315.89’’
analysis, e.g. FEM or FDM software, via the established response if trial B = 4, and either B value can be used. (The need to
surface method, artificial neural network method, or other similar start the search from initial values which result in positive
methods. g(x) value is to ensure that the point from which search
Key features of the spreadsheet FORM approach as illustrated in starts is in the safe domain. This is further illustrated in Sec-
Fig. 3 are: tion 8 in connection with reliability-based design of an
anchored sheet pile wall.)
(i) The boxed cells in rows 9 and 11–14 (except cells B14 and (iv) Microsoft Excel Solver is invoked, to Minimize (automati-
D14) contain established equations related to bearing capac- cally) cell D14, By Changing (automatically) cells G3:G6, Sub-
ity, including Eqs. (6b,c) and (7a,b,c). ject to the Constraints cell B14 = 0. If cell B9 for foundation
(ii) Cell B14 contains Eq. (6a), and cell D14 contains the array width is 3 m, a reliability index b (cell D14) of 0.11 will be
formula ‘‘ = SQRT(MMULT(TRANSPOSE(G3:G6), MMULT obtained by Solver. This falls short of the required b of 3.0.
(MINVERSE(matrix R),G3:G6)))’’, which computes reliability The cell B9 value is increased from 3 m and Solver is invoked
index b by Eq. (3) via a constrained optimization program again. After a few trials a foundation width of 4.51 m is
Solver which is part of Microsoft Excel. found to yield a reliability index b of 3.0, as shown in Fig. 3.
(iii) The initial values for the column labeled n were zeros. Input
initial trial value in cell B9 (for foundation width B) such that As mentioned in the Solver Parameters window (the screen-cap-
the Performance Function g(x) in cell B14 displays a positive ture shown in the lower part of Fig. 3), the Solver’s default solving
computed value when cells G3:G6 are zeros initially. For the method is the Generalized Reduced Gradient (GRG) Nonlinear
34 B.K. Low, K.-K. Phoon / Computers and Geotechnics 65 (2015) 30–44

engine for Solver Problems that are smooth nonlinear. In the 2010 parameters are multiplied by partial factors to obtain the design
version of Microsoft Office, the ‘‘Evolutionary engine’’ can also be values. The resistance thus diminished is required to be greater
selected for Solver problems that are non-smooth. Also, although or equal to the amplified actions (loadings) as shown in Fig. 4,
only 4 changing variables (cells G3:G6) and 1 constraint (that cell which also summarizes the three design approaches DA1, DA2
B14 = 0) are specified for the problem in hand, in principle up to and DA3. The DA2 is akin in principle to the load and resistance
200 variable cells and numerous constraints of different forms factor design (LRFD) of North America. With respect to characteris-
can be specified. tic value, Clause 2.4.5.2(2)P [16] defines it as being ‘selected as a
If the need arises for verifying whether the real minimum has cautious estimate of the value affecting the occurrence of the limit
been obtained according to Eq. (3) (and hence Eq. (2) or (1)), one state’, and Clause 2.4.5.2(10) of EC7 states that statistical methods
can re-invoke Solver after initializing the column labeled n (cells may be used when selecting characteristic values of geotechnical
G3:G6) to some randomized values, and check whether the same parameters, but they are not mandatory.
solution is obtained. For the case in hand, 10 Solver searches (step One may note the following similarities and differences
(iv) above) each with randomized initial n values between 1 and between a reliability-based design and that based on EC7, in the
+1 yielded identical solutions. (Randomized initial n values are context of the reliability-based design of Fig. 3:
input readily using Microsoft Excel’s random number generation
tool, for example by specifying random n values between 1 and (a) The x⁄ column denotes the point where the 4-dimensional
+1based on uniform distribution.) equivalent dispersion ellipsoid touches the limit state sur-
The probability of failure (PerFunc < 0, or qu < q, Eq. (6)) inferred face. It is the most probable failure combination of the
from the reliability index b = 3.0 using Eq. (5) is 0.135%. This com- parameters.
pares very well with the results of six Monte Carlo simulations (b) The mean value, standard deviation and design point value
each with 800,000 iterations from Latin Hypercube sampling using of a parameter can be related to the characteristic value
the commercial software @RISK (http://www.palisade.com), which and partial factor of the same parameter in the EC7 design
yielded six failure probabilities of 0.117%, 0.125%, 0.134%, 0.126%, approach. For example, for the given mean and standard
0.121% and 0.123%. The mean of the six Monte-Carlo estimated Pf deviation of the angle of friction /, the characteristic value
is 0.124%. In Monte Carlo simulations, the 95% confidence interval / (assuming at 10 percentile of the lognormal distribution)
for the probability of unsatisfactory performance can be estimated is 21.9°. The partial factor of / (denoted by the symbol c/)
from the following equation [41]: implied in the design point (the x⁄ values) is
sffiffiffiffiffiffiffiffiffiffiffiffiffiffi c/ = /k//⁄ = 21.9°/20.77° = 1.054. (Note that in EC7 the par-
1  Pf tial factor c/ is applied to tan /k, not to /k. If the mean and
%error ¼ 200 ð8Þ
nPf standard deviation are given for tan / instead of /, the char-
acteristic value tan /k and its corresponding partial factor
where n is the total number of simulation trials. The range of scatter can be back-calculated from the FORM results by the same
of estimated Pf in the six Monte Carlo simulations (between 0.117% principle. Partial factors thus back-calculated depend on
and 0.134%) is consistent with Eq. (8) which indicates a %error of assumed characteristic values.)
6.4%, based on Pf  0.12% and n = 800,000. Hence, the 95% confi- (c) On the other hand, for the horizontal load parameter Qh,
dence interval for the probability of unsatisfactory performance Pf with mean 300 kN/m and standard deviation 45 kN/m, the
is, approximately, (10.064)  .124% 6 Pf 6 (1+0.064)  .124%, or characteristic value Qhk (assuming at 90 percentile of the
0.116% 6 Pf 6 0.132%. The preceding numerical discussions were lognormal distribution) is 359 kN/m. The partial factor of
meant to emphasize that Monte Carlo simulations, even with itera- Qh (denoted by the symbol cQh) implied in the design point
tions close to 1 million, does not produce precise estimate of failure (the x⁄ values) is cQh = Qh⁄/Qhk = 412.6/359 = 1.15.
probability when the failure probability is of the order of 0.1% (d) For lower mean value and standard deviation of Qh, or for
which is implied by a design based on a reliability index b of 3.0. different applications (e.g., slope stability or earth retaining
More refined sampling (e.g., the importance sampling method with walls), the back-calculated partial factors could be different.
sampling near the FORM design point) can narrow (but not elimi- To apply the same specified partial factors of EC7 across dif-
nate) the scatter of estimated failure probabilities from multiple ferent applications or different levels of parametric uncer-
sessions of simulations. tainty may not imply the same target failure probability. In
If desired, the original correlation matrix can be modified in line contrast, in a reliability-based design, the same target reli-
with the equivalent normal transformation, as suggested in Der ability index (with its implied probability of failure) can be
Kiureghian and Liu [13]. For the cases illustrated herein, the corre- used across different applications and different levels of
lation matrix thus modified differs only very slightly from the ori- parametric uncertainty and correlations, and including
ginal correlation matrix. Hence, for simplicity, the examples of this asymmetric probability distribution if appropriate. Also, if
study retain the original unmodified correlation matrices. Alterna- higher reliability is required where the consequence of fail-
tively, instead of matching the Pearson correlation coefficient ure is more significant, a safer design can be obtained by
which forms the basis of Der Kiureghian and Liu [13], one can raising the target reliability index. Such flexibility and auto-
adopt the approach of matching the Spearman rank correlation. matic incorporation of parametric sensitivities and correla-
It is not necessary to modify the correlation matrix for non-normal tions into the design process are not found in a Eurocode
random variables in this case. Details are given by Li et al. [21]. design based on code-recommended partial factors.
The similarities and differences between this reliability-based (e) Table B2 in EN1990:2002 recommends the minimum reli-
design and design based on EC7 are discussed next. ability index for three reliability classes (RCs) for the ulti-
mate limit state. For RC1, RC2, and RC3, the minimum
4. Reliability-based design (RBD) compared with Eurocode 7 reliability indices for a 50 year reference period are 3.3,
(EC7), and complementary roles of RBD to EC7 design 3.8, and 4.3, respectively. Eurocode 7 describes the concept
of ‘‘Geotechnical Category’’ in Section 2.1 ‘‘Design’’ require-
In geotechnical design based on EC7, the characteristic values of ments, but there is no statement linking a geotechnical cat-
resistance parameters are divided by partial factors to obtain the egory to a reliability class despite an implied intent relating
design values, while the characteristic values of action (load) to reliability differentiation and no guideline on how to
B.K. Low, K.-K. Phoon / Computers and Geotechnics 65 (2015) 30–44 35

General concepts of ultimate limit state design in Eurocode 7:


(if “=“, then “design point”)

Diminished resistance (ck / γc, tanφk / γφ) > Amplified loadings

Characteristic values Based on characteristic values and


partial factors for loading
Partial factors parameters.

“Conservative”, for example, 10 percentile for strength parameters, 90


percentile for loading parameters

The three sets of partial factors (on resistance, actions, and material properties) are
not necessarily all applied at the same time.
In EC7, there are three possible design approaches:
● Design Approach 1 (DA1): (a) factoring actions only; (b) factoring materials only.
● Design Approach 2 (DA2): factoring actions and resistances (but not materials).
● Design Approach 3 (DA3): factoring structural actions only (geotechnical actions
from the soil are unfactored) and materials.

Fig. 4. Characteristic values, partial factors, design point, and design approaches (DA) in Eurocode 7.

modify partial factors or characteristic values to suit differ- (i) The x⁄ value for c is slightly higher than the original mean
ent geotechnical categories. Based on Section B3.3 and value of c, due to the negative correlation between c and /
Table B3 in EN1990:2002, it would appear that partial fac- . In this case, the response is far more sensitive to / than
tors in all Eurocodes are meant for conventional RC2 and to c. Negatively correlated c and / means low values of /
reliability differentiation is achieved by reducing load fac- tend to occur with high values of c, and vice versa. Under
tors by 10% for RC1 and increasing load factors by 10% for such circumstances, a reliability-based design will automat-
RC3. Partial factors associated with resistance are left ically reflect sensitivities in obtaining design point values of
unchanged. Given that Eurocode 7 is intended to be used c and / which can be both lower than their respective mean
in conjunction with EN1990:2002 as stated in Section 1.1.1 values (as in the reinforced rock slope of the next section), or
‘‘Scope of EN1997’’, one may infer that the same design one above average and the other below average as in this
approach is applicable when a geotechnical category 1 or 3 case. For the case in hand, to have both design values of c
situation is encountered. Reliability-based design can easily and / below their mean values using partial factors is unre-
accommodate different target reliability indices such as alistic. This lack of realism arises because negative correla-
those recommended in Table B2 of EN1900:2002 and it tion between c and / cannot be considered in EC7 and it is
can thus play a useful role in validating the 10% load factor difficult for engineers to decide how to factor up c and factor
guideline for non-conventional design scenarios that cannot down / in a coordinated manner consistent with their neg-
be classified within geotechnical category 2. ative correlation relationship by pure intuition or judgment.
(f) Note that the back-calculated partial factors are corollary In the presence of a correlation matrix reflecting all possible
by-products of a reliability-based design, which requires sta- correlations between multiple parameters, it will be impos-
tistical inputs (mean values, standard deviations, parametric sible to factor up/down correctly without reliability analysis.
correlations, probability distributions) but not partial factors
or characteristic values. The design point is obtained as the The above discussions suggest that reliability-based design can
most probable failure combinations of parametric values, provide additional insights to EC7 design or LRFD design when the
and the reliability index b, being the distance (in units of statistical information (mean values, standard deviations, correla-
directional standard deviations) from the safe mean-value tions, probability distributions) of the key parameters affecting
point to the nearest failure boundary (the limit state surface, design are known and one or more of the following circumstances
Fig. 1), conveys information on the probability of failure. In apply:
Eurocode design or LRFD, there is no explicit information
on the probability of failure.  When partial factors have yet to be proposed by Eurocode 7 to
(g) It is clear from the above discussions that in a reliability- cover uncertainties of less common parameters, for example,
based design one does not use or specify the partial factors. in situ stress coefficient K in underground excavations in rocks,
The design point values x⁄ (and implied characteristic values dip direction and dip angles of rock discontinuity planes, and
and partial factors) are determined automatically and reflect the smear effect of vertical drains installed in soft clay.
sensitivities, standard deviations, correlation structure, and  When output has different sensitivity to an input parameter
probability distributions in a way that prescribed partial fac- depending on the engineering problems (e.g. shear strength
tors and ‘‘conservative’’ characteristic values cannot. parameters in bearing capacity, retaining walls, slope stability).
(h) In Fig. 3, by comparing the n values in cells G3:G6, it is evi-  When input parameters are by their physical nature either pos-
dent that bearing capacity is more sensitive to Qh than to Qv, itively or negatively correlated. One may note that EC7 does not
for the case in hand with its statistical inputs. mention use of different characteristic values and partial factors
36 B.K. Low, K.-K. Phoon / Computers and Geotechnics 65 (2015) 30–44

when some parameters are correlated. The same design is with water (as characterized by the ratio zw/z) are also negatively
obtained in EC7 with or without modeling of correlations correlated. This means that shallower crack depths tend to be
among parameters. water-filled more readily (i.e., zw/z ratio will be higher) than deeper
 When spatially autocorrelated soil properties need to be mod- crack depths, consistent with the scenario suggested in Hoek [20]
eled. This will be illustrated in Section 6 on a Norwegian slope that the water which would fill the tension crack in this Hong Kong
with spatially autocorrelated soil unit weight and undrained slope would come from direct surface run-off during heavy rains.
shear strength, and in Section 7 on a laterally loaded pile. (On the other hand, positive correlation between z and zw/z may
 When there is a target reliability index or probability of failure be justified if the water depth zw in Fig. 5 is governed by a prevail-
for the design in hand. The target reliability index or probability ing water table.) For illustrative purposes, a negative correlation
of failure can be set explicitly higher or lower in RBD depending coefficient of 0.5 is assumed between z and zw/z, as shown in
on the importance of the structure or consequence of failure. In entries R43 and R34, where Rij denotes entry in row i and column j
this regard, one may note that a design by EC7 or LRFD provides of the correlation matrix R in Fig. 5. Even though there are no data
no explicit information on the reliability level or probability of to quantify this correlation between z and zw/z, it is still useful to
failure, nor explicit guidance on how characteristic values or explore possible correlations to get a feel for its influence on the
partial factors are to be modified when more stringent safety reliability index. This is a sensible approach commonly applied in
requirements are needed. engineering practice for important but not well characterized
 When uncertainty in unit weight c of soil needs to be modeled. parameters.
This will be illustrated in Section 6 on a Norwegian clay slope, Fig. 5 shows that a reinforcing force T of 257 tons (per m length
and in Section 8 on the reliability-based design of an anchored of slope) inclined at h = 55° (measured from the perpendicular to
sheet pile wall. the sliding surface) is needed to achieve a target reliability index
of 3.0. The most sensitive parameters for the case in hand are the
The following sections illustrate geotechnical reliability-based ratio zw/z and the coefficient of horizontal earthquake acceleration
design and analysis, focusing on subtleties and insights, and dis- a. Note that the design point values of resistant parameters c and /,
cussions in the context of the EC7 design approach. The cases at 8.11 t/m2 and 29.65° respectively, are lower than their mean val-
involve the ultimate limit states of a reinforced rock slope, a soil ues of 10 t/m2 and 35° respectively. This is in contrast to the case of
slope with spatial variability, an anchored sheet pile wall, and a shallow footing subjected to an inclined load (Fig. 3), where the
the serviceability limit state of a laterally loaded pile. The comple- design point of c, at 15.2 kPa, is slightly above its mean value of
mentary roles that RBD can play toward EC7 design and RBD’s 15 kPa, while the design point value of /, at 20.77°, is below its
potential to enhance EC7 design will be emphasized. mean value of 25°. The contrasting locations of the design points
The Low and Tang [30] spreadsheet-based x-via-n FORM proce- are explained next.
dure will be used throughout. The other two methods outlined in For the reinforced rock slope of Fig. 5, the design point is where
Fig. 2, namely the [26] FORM method (directly varying x) and the the expanding ellipsoid (or equivalent dispersion ellipsoid when
x-via-n-via-u FORM method (a slight variation of the [30] method), nonnormal distributions are involved) is tangent to the limit state
will obtain the same reliability index and design point for all the surface at a point where c < lc and / < l/, as shown in Fig. 1 for a
examples in this study. case with two random variables (c and /) only. Had the expanding
ellipsoid in Fig. 1 touched the limit sate surface at the upper left
portion (possible if the limit state surface has a steeper negative
5. Reliability-based design of a reinforced rock slope gradient which implies greater sensitivity to the angle of friction
/ than to the cohesion c), the design point value of c would be
Hoek [20] presents a limit equilibrium formulation for a 2- higher than its mean (lc) while that of / would be lower than its
dimensional rock slope in Hong Kong, leading to the factor of safety mean (l/), as is the case in the shallow footing subjected to
Fs against sliding along a joint plane. The limit equilibrium formu- inclined load (Fig. 3), and vice versa if the tangency point is at
lation of Hoek [20] was used in a probabilistic rock slope analysis the lower right part of the limit state surface (possible if the limit
in Low [29], in which the reliability index b was computed using state surface has a very gentle negative gradient which implies
the Low and Tang [26] FORM procedure (Fig. 2, top left), and the greater sensitivity to the cohesion c than to the friction angle /).
effects of statistical correlations on the computed reliability index One may also appreciate that positively correlated resistance
were also investigated. This Hong Kong slope is further investi- parameters (implying that low resistances tend to occur together)
gated in this section with respect to reliability-based design using will have lower reliability index than uncorrelated resistance
the Low and Tang [30] FORM procedure (Fig. 2, top right). It will be parameters, as shown in Figs. 2 and 4 of Low [27]. A reliability-
shown that a more precise estimate of failure probability is achiev- based approach like the one presented here is able to locate the
able by performing second-order reliability method (SORM) based design point case by case and in the process reflect parametric sen-
on the design point and reliability index obtained in FORM. A prag- sitivities (related to the limit state surface and hence application
matic stand on the sufficiency of FORM for reliability-based design specific) and correlation structure in a way that design based on
will be proposed. prescribed partial factors cannot.
Fig. 5 shows the FORM reliability-based design of the 2-dimen- In the retaining wall foundation problem of Fig. 3, the bearing
sional rock slope with five correlated random variables, two of capacity factors Nc, Nq and Nc are functions (Eq. (7)) of the angle
which obey the highly asymmetric truncated exponentials. The of friction /. In addition, the three load inclination coefficients iq,
five random variables are: the shear strength parameters c and / ic and ic in Fig. 3 are (as formulated according to [3]) also functions
of the failure surface (the joint plane with inclination angle wp), of the angle of friction /. Hence the bearing capacity under the
tension crack depth z, height of water zw in the tension crack, given inclined loading is much more sensitive to / than to c, for
and the coefficient of horizontal earthquake acceleration a. the parametric uncertainties and correlations shown. In contrast,
The negative correlation coefficient of 0.5 between c and / the rock slope stability of Fig. 5 is about equally sensitive to both
shown in the top left of the correlation matrix R in Fig. 5 means the cohesion c and the friction angle / of the rock joint, as judged
that low values of cohesion c tend to occur with high values of fric- from their respective n values of 0.947 and 1.069.
tion angle /, and vice versa. In addition, one may logically infer For the rock slope of Fig. 5, when there is no reinforcing force,
that the tension crack depth z and the extent to which it is filled i.e., T = 0, the reliability index obtained by Excel Solver is
B.K. Low, K.-K. Phoon / Computers and Geotechnics 65 (2015) 30–44 37

Units: meter, tonne,


tonne/m2, tonne/m3.
Tension
crack

FORM FORM Monte Carlo


αW followed by simulations,
V Z SORM, sample size
W Zw
Sample size = 8 800,000
H θ U for estimating
Assumed curvature
water
Failure
ψf Tpressure probability
0.13% 0.060% 0.062%
ψp Failure surface

x* Para1 Para2 Para3 Para4 n Correlation matrix R u


Normal c 8.106 10 2 -0.947 1 -0.5 0 0 0 -0.947
Normal φ 29.653 35 5 -1.069 -0.5 1 0 0 0 -1.782
Normal z 13.831 14 3 -0.056 0 0 1 -0.5 0 -0.056
Tr_Exp zw /z 0.801 0.5 0 1 1.428 0 0 -0.5 1 0 1.616
Tr_Exp α 0.133 0.08 0 0.16 1.527 0 0 0 0 1 1.527

H ψf ψp γ γw T θ g(x) β Pf(FORM)
60 50 35.0 2.6 1 257 55 0.000 3.00 0.0013

Fig. 5. A horizontal reinforcing force of 257 tons/m is required to provide the rock slope with a reliability index b of 3.0. The solution values in the column labeled n were
obtained automatically using Excel Solver.

b = 1.887, implying a probability of failure of 2.96% by Eq. (5), components of curvature at the design point in the 5-dimensional
which is unacceptably high. A few trials of the value of T, each random variable u-space are based on sampling grid coefficient
invoking Excel Solver to minimize the b cell by automatically k = 1 if b 6 3, and k = 3/b if b > 3. The accuracy of this FORM-
changing the n vector (initially zeros) subject to the constraint that then-SORM Pf estimation is confirmed by six Monte Carlo simula-
the g(x) cell be equal to 0 (similar to step (iv) in Section 3), lead to tions each with 800,000 Latin Hypercube sampling, yielding
the solution as shown, namely a horizontal (for which h = 55°) T 0.063%, 0.062%, 0.068%, 0.059%, 0.061% and 0.061%, or an average
force of 257 tons/m is required to achieve a reliability index b of failure probability of 0.062%, practically the same as the Pf(SORM)
3.0, corresponding to a failure probability of about 0.13% by Eq. (5). value of 0.06%. The number of sampling points (eight) for estimat-
Eq. (5) is exact when the LSS is planar and the parameters fol- ing curvature at the FORM design point is five orders of magnitude
low normal distributions. Inaccuracies in Pf estimation may arise smaller than the 800,000 in a Monte Carlo simulation.
when the LSS is curved and/or nonnormal distributions are mod- The above RBD of a reinforced rock slope and the RBD of a
eled. More refined alternatives are available, for example the retaining wall foundation (Section 3) illustrate the complementary
established second-order reliability method (SORM). SORM analy- roles of RBD as described in bullet points 1, 2, 3 and 5 at the end of
sis requires the FORM b value and design point values as inputs, Section 4, namely that RBD can be used when there are no EC7 rec-
and therefore is an extension dependent on FORM results. In gen- ommended partial factors yet (e.g., the parameter zw/z in Fig. 5),
eral, the SORM attempts to assess the curvatures of the LSS near when the engineering problems (Figs. 3 and 5) show different sen-
the FORM design point in the dimensionless and rotated u-space. sitivities to the same parameters c and / which cannot be reflected
The failure probability is calculated from the FORM reliability by fixed partial factors, when the parameters are statistically cor-
index b and estimated principal curvatures of the LSS using estab- related based on physical considerations (e.g. between c and /,
lished SORM equations of the following form: and between z and zw/z in Fig. 5), and when there is a target reli-
ability index and associated failure probability which can be higher
Pf ðSORMÞ ¼ f ðbFORM ; Curvatures at the FORM design pointÞ ð9Þ
or lower depending on the consequence of non-performance. More
These SORM formulas have been used by Chan and Low [4], who illustrations of how RBD complements EC7 will be provided in the
presented a practical and efficient spreadsheet automated sections to follow. But first some pragmatic discussions on FORM
approach of implementing SORM using an approximating parabo- and SORM are necessary and are done in the next paragraph.
loid [14] fitted to the LSS in the neighborhood of the FORM design Engineers can adopt one of the following positions in perform-
point. Complex mathematical operations associated with Cholesky ing reliability-based design:
factorization, Gram-Schmidt orthogonalization and inverse trans- Pragmatic Approach A, using FORM only: Conduct RBD based
formation are relegated to relatively simple short function codes on FORM, for a target reliability index of 3.0 (or higher depending
in the Microsoft Excel spreadsheet platform. on importance of structure and consequence of failure). An approx-
The Pf(SORM) value of 0.06% shown in the table at the top right imate failure probability Pf can be estimated from Eq. (5), with the
of Fig. 5 was obtained using the Chan and Low [4] Excel spread- pragmatic stand that although the correct Pf can differ from the
sheet SORM approach. The 8 sample points for estimating the 4 approximate Pf, both are sufficiently small (e.g., confident to a high
38 B.K. Low, K.-K. Phoon / Computers and Geotechnics 65 (2015) 30–44

degree that Pf < 0.5% when b = 3.0, and Pf < 0.1% when b = 3.5). This geomaterial with some unique statistical features (e.g. spatial
is in line with the spirit of EC7 design approach which aims at autocorrelation) not commonly found in structural material manu-
achieving a sufficiently safe design, not at a specific value of low Pf. factured under strict quality control. For example, by the nature of
Rigorous Approach B, using FORM and SORM: Conduct RBD the slow precipitation (over many seasons) of fine-grained soil par-
based on FORM for a target reliability index value (e.g. 3 or 3.5), ticles in water in nearly horizontal layers, two points in close hor-
followed by SORM for a more accurate estimate of failure probabil- izontal or vertical proximity to one another are likely to be more
ity Pf. If desired, adjust the design based on SORM Pf so as to positively correlated (likely to have similar undrained shear
achieve the desired target Pf, as explained later. strength cu values, for example) than two points further apart in
The authors recommend the Pragmatic Approach A to practitio- the vertical direction. This is manifested in some clays such as var-
ners new to RBD via FORM. The FORM-followed-by-SORM RBD Rig- ved clays.
orous Approach B can be adopted, if desired, when there is sufficient A clay slope in southern Norway was analyzed deterministically
familiarity with the approach in the target-b-via-FORM of Prag- and probabilistically by Low et al. [31] using the Low [25] spread-
matic Approach A. Both the Pragmatic Approach A and the Rigorous sheet-based reformulations of Spencer method and the intuitive
Approach B do not involve Monte Carlo simulations, and for prob- first-order reliability method of Low and Tang [26]. The reformula-
lems without closed-form solutions, established methods like the tion allows switching among the Spencer, Bishop simplified and
response surface method or neural network approach can be used wedge methods on the same template, by specifying different
to bridge the spreadsheet FORM approach with stand-alone FEM or side-force inclination options and different constraints of optimi-
FDM programs. zation. Search for the critical circular or noncircular slip surface
If the Rigorous Approach B is desired, a practical approach for is possible. The deterministic procedure was extended probabilisti-
revising the design is as follows. In the light of the comparison of cally by implementing the first-order reliability method via con-
FORM Pf of 0.13% vs. actual Pf of 0.060% in Fig. 5, suppose a proba- strained optimization of the equivalent dispersion ellipsoid in the
bility of failure of 0.1% is the target Pf for the case in hand, one may original space of the random variables. The procedure was illus-
do a FORM design of the reinforcing force T for a reliability index trated for an embankment on soft ground, and for a clay slope in
corresponding to a probability of failure (by Eq. (5)) of (0.13%/ southern Norway, both involving spatially correlated soil proper-
0.060%) ⁄ 0.1% = 0.22%, i.e. b ¼ U1 ð1  0:0022Þ ¼ Excel function ties. The effects of autocorrelation distance on the results of reli-
NormSInv(0.9978) = 2.85. For the case in Fig. 5, the required T force ability analysis were studied. Shear strength anisotropy was
is then found (via Excel Solver runs) to be 224 tons/m for a target modeled via user-created simple function codes in the program-
b = 2.85. This downward revision of T force from the original ming environment of the spreadsheet.
257 tons/m takes about a minute. To verify, with input T = 224, Fig. 6 shows the results of reliability analysis involving 24 spa-
four Monte Carlo simulations each of 800,000 random sets yielded tially correlated cu values and 24 spatially correlated unit weight
probabilities of failure of 0.114%, 0.100%, 0.112% and 0.101% values. The lognormal distribution is assumed for random vari-
respectively; the average is practically the target Pf of 0.1%. ables. The cu values and the unit weight c values of the 24 slices
The above reliability-based design procedure (no Monte Carlo form two horizontally auto-correlated series based on the estab-
simulations needed) may be summarized as follows [34]: lished exponential autocorrelation function:
jxmid;i xmid;j j
(i) Conduct reliability-based design based on initial target reli- qij ¼ e d ð11Þ
ability index bi of 3.0 or 3.5. The implied FORM Pf is U(bi),
equal to 0.13% and 0.023%, respectively. where xmid,i and xmid,j are the x-coordinates of the mid-points of the
(ii) For the design solution of (i), obtain more accurate Pf using base of slice i and slice j respectively.
SORM on the foundation of FORM design point and b. This The size of the correlation matrix is 48  48. The design point
Pf is denoted as Pf,SORM. obtained by Excel Solver represents the most probable combina-
(iii) Decide on a target Pf (<1%), and repeat the reliability-based tion of the 24 values of cu and the 24 values of c which would cause
design based on an updated target reliability index (brev) as failure. As expected for resistance parameters, the 24 values of
follows: undrained shear strength cu at the design point are all lower than
their respective mean values. On the other hand, when the auto-
  correlation distance d is 10 m or lower, as shown in Fig. 6, the
Uðbi Þ
brev ¼ U1 1   target Pf ð10Þ design point index of c—defined as (ci⁄  lc)/rc where ci⁄ is the
P f ;SORM
design point value of unit weight c for i from 1 to 24—shows most
The final design solution is obtained via repeated FORM analy- values of unit weight c⁄ above their mean value lc, as expected for
ses using different designs (e.g. foundation width B) until the reli-
ability index is equal to brev. Linear interpolation between two
designs and their respective reliability indices can also be done. 20
Elevation (m)

The final design based on brev will have failure probability close 10
to the target Pf. − 0.6 m
0
If the design is to be obtained by direct trial and error of differ- 70 90 110 130 150 170 190
ent designs each checked against Monte Carlo simulations (with or -10

without importance sampling), the computation time will be very -20


much longer than the FORM-and-SORM approach involving Eq.
1.5
(10). 1
δ = 1000
Design Point

0.5
Index of γ

100
0 50
6. Reliability analysis of a Norwegian slope accounting for -0.5 70
10 110 150 190
5
spatial autocorrelation -1 2
-1.5 x-Coordinates (m)
1
-2
Spatial autocorrelation (also termed spatial variability) arises in
geological material by virtue of its formation by natural processes Fig. 6. Design point index of clay unit weight as a function of horizontal
acting over geologic time (order of millions of years). This endows autocorrelation distance d in meter. [31].
B.K. Low, K.-K. Phoon / Computers and Geotechnics 65 (2015) 30–44 39

loading parameters, but, interestingly, there are some design-point curves, Fig. 7. The soil–pile interaction problem was solved using
values of c near the toe which are below their mean values. The a rigorous numerical procedure based on constrained optimization
implication is that the slope is less safe when the unit weights near in the spreadsheet platform. The numerical procedure was then
the toes are lower. This implication can be verified by deterministic extended to reliability analysis in which the Hasofer–Lind index
runs using higher c values near the toe, with resulting higher fac- was computed. The soil resistance was modeled stochastically to
tors of safety. It would be difficult for Eurocode 7 to recommend reflect spatial variation. Multicriteria reliability-based design was
partial factors such that the design values of c are above the mean also illustrated. The work was further extended in Chan and Low
along some portions of the slip surface and below their mean along [5] to a probabilistic analysis of laterally loaded piles using
other portions. In contrast, the design point is located automati- response surface and neural network approaches. A case of deter-
cally in FORM analysis by the spreadsheet’s built-in constrained ministic analysis followed by multi-criteria reliability-based
optimization program, and reflects sensitivity and the underlying design from Low et al. [24] is presented in this section, with new
statistical assumptions from case to case in a way specified partial discussions on how RBD complements EC7 design.
factors cannot. The deterministic problem shown in the top right of Fig. 7a is
At higher values of autocorrelation distance d in Eq. (11), the described in Tomlinson [43], which used manual computation to
correlation coefficients approach 1.0; the design point indices of compute approximate pile head deflection whereas the solution
c of the 24 slices approach a common value, as shown by the nearly in Fig. 7a is obtained using the rigorous numerical procedure
horizontal line in Fig. 6 for d = 1000 m. The design point indices of described in Low et al. [24] and Chan and Low [5]. The steel tubular
cu—defined as (cui⁄  lcui)/rcui—of the 24 slices also approach a uni- pile in Fig. 7a has an outside diameter d of 1.3 m and a wall thick-
form common value when d = 1000 m; however, the individual ness of 0.03 m and is part of a pile group in a breasting dolphin. The
design-point values of cu differ from slice to slice because the flexural rigidity EpIp of the pile is 4,829,082 kNm2. The pile is
mean, lcui, and standard deviation, rcui, vary from slice to slice. embedded 23 m in the stiff overconsolidated clay with average
The implications of not considering seabed erosion versus treat- undrained shear strength cu = 150 kN/m2 near sea bed, and pro-
ing seabed as random (to account for uncertain depth of erosion) trudes 26 m above the sea bed. For the case where a cyclic force
were discussed by Low et al. [31]. of 421 kN is applied at 26 m above the seabed, only the embedded
The results of reliability analysis are only as good as the statis- portion of the pile requires soil-structure interaction analysis. The
tical inputs and reliability method used (e.g., FORM or SORM), in deflection of the pile head at 26 m above the seabed can be inferred
the same way that the results of deterministic analysis are only from statics once the deflection and rotation of the pile at seabed
as good as the deterministic inputs and method used (e.g. Spencer level are known. The Matlock p–y curves for clays, as shown in
method or other methods). A reliability analysis requires addi- Fig. 7b, have been used. The p–y curves are nonlinear, exhibit
tional statistical input information, which is not required in a strain-softening, and vary with depth. The calculated pile deflec-
deterministic factor-of-safety approach but results in richer infor- tion (yi) at seabed level (where z = 0) is 0.0602 m. The pile head
mation pertaining to the performance function and the design deflection (at 26 m above the seabed) is, by integrating the
point that is missed in a deterministic analysis. moment-curvature equation, 0.994 m or about 1 m.
The FORM reliability approach reflects the underlying analytical For reliability analysis, the 23 m embedded length of the pile
formulations and statistical assumptions and is able to locate the was discretized into 30 segments of progressively greater length.
most probable combination of parametric values which would The random variables were the lateral load PH at pile head and
cause failure and the corresponding reliability index, without rely- the undrained shear strength cu at 31 nodal points along the
ing on rigid partial factors. embedded pile length below the seabed. The PH was assumed to
It may be noted that all partial factor approaches including be normally distributed, with mean value 421 kN and a coefficient
Eurocode 7 cannot address a system reliability problem effectively. of variation of 25%. Since the mean undrained shear strength cu
Slopes with multiple soil layers may involve multiple failure typically exhibits an increasing trend with depth, it was assumed
modes and they are fundamentally system reliability problems. that the mean trend is lcu = 150 + 2z, kPa. The standard deviations
Details are reported by Ching et al. [6], Ching et al. [7]. System reli- of the 31 cu random variables were equal to 30% of their respective
ability analysis via FORM for a multi-layer slope with multiple fail- mean values. The following established negative exponential
ure modes is presented by Low et al. [33]. model was adopted to model the vertical spatial variation of the
Spatial variability of soil properties is considered also in the cu values:
next RBD example, but logically in the vertical direction of a later-
ally loaded vertical pile (Eq. (12) in the next section) instead of the qij ¼ e
jDepthðiÞDepthðjÞj
d ð12Þ
horizontal spatial variability modeled by Eq. (11). It will be shown
that the sensitivity of the pile head lateral deflection to the spa- The b index obtained was 1.514 with respect to yielding at the
tially autocorrelated undrained shear strength cu values depends outer edge of the annular steel cross section.
on the protruding length of the pile above the seabed. RBD auto- At the most probable failure point (where the hyperellipsoid
matically reflects such different cu sensitivities from case to case touches the limit state surface) the value of the lateral load at pile
with different protruding pile lengths above seabed, and thus can head is PH = 580.3 kN, i.e., at 1.513 rPH from the mean value of PH,
play a very useful complementary role to the Eurocode 7, which while the 31 autocorrelated cu values deviates only very slightly
has no provision yet for spatially autocorrelated soil properties or from their mean values. This is not surprising given the e = 26 m
for different partial factors on cu depending on the protruding pile cantilever length above the seabed; in fact the maximum bending
lengths above seabed. moment occurs at a depth of only 1.36 m below the seabed, or
27.36 m from the pile head. Hence, for the case in hand, pile yield-
ing caused by bending moment is sensitive to the applied load at
7. Multi-criteria reliability-based design of a laterally loaded pile head, and not sensitive to the uncertainty of the shear strength
pile in clay with spatially autocorrelated shear strength below seabed. However, separate reliability analysis for cases
where the lateral load acts on pile head near ground surface (with
Low et al. [24] studied the deflection and bending moment of zero cantilever length) indicates that the response is sensitive to
laterally loaded single piles in which soil-pile interaction was both the lateral load at pile head and the soil shearing resistance
based on the nonlinear and strain-softening Matlock [37] p–y within the first few meters of the ground surface. The different
40 B.K. Low, K.-K. Phoon / Computers and Geotechnics 65 (2015) 30–44

Deflection y (m) Moment M (kNm) Soil reaction (kN/m) PH = 421 kN

-20000 0 20000
y=1m
-0.05 0.00 0.05 0.10 -500 0 500
0 0 0
seabed seabed
e = 26 m
5 5 5
water

10 10 10 y = 60 mm
z=0
seabed
15 15 15

20 20 20 Stiff clay steel pipe pile,


cu = 150 kPa d=1.3m
25 25 25
z = 23 m
(a)

1500
z = 22 m
γ' = 11.8 kN/m3
1000 c u= 150 kN/m 2
J = 0.25 12 m
B = 1.3 m
500 ε = 0.01
1.5
p (kN/m)

z=0m
External diameter (m)

β = 3, (bending)
0 c
z=0m 1.45
-500 β = 3, (def lec tion)
a
1.4
-1000

z = 22 m b
-1500 1.35
-0.3 - 0.1 0.1 0.3 30 32 34 36 38 40
y (m) Steel wall thickness (mm)

(b) (c)
Fig. 7. Multi-criteria reliability-based design of a laterally loaded pile in spatially autocorrelated clay: (a) numerical nonlinear p–y analysis of a steel tubular pile in a breasting
dolphin, (b) Matlock’s nonlinear p–y curves, (c) combinations of pile external diameter and wall thickness for reliability index b of 3. [24].

sensitivities from case to case is automatically reflected in reliabil- ministic case for which an average cu value of 150 kPa was used
ity analysis aiming at a target index value (presented next), but as in Tomlinson [43]). This average deflection of 0.986 m does
will be difficult to consider in codes based on partial factors. The not provide information on the reliability of not exceeding the
useful insights offered by RBD to complement EC7 design are obvi- 1.4 m tolerable limit, because the uncertainties of the random vari-
ous, because RBD automatically reflects such different cu sensitivi- ables have not been reflected in estimates based on mean values.
ties from case to case with different protruding pile lengths above To illustrate multi-criteria reliability based design, suppose it is
seabed. desired to select external diameter d and steel wall thickness t of
the tubular pile so as to achieve a reliability index of 3.0 with
7.1. Illustrative example of multi-criteria reliability-based design of a respect to both the pile head deflection limit state and pile bending
laterally loaded pile moment limit state. The mean and covariance structure of PH and
cu profile are as described above. Pile embedment length is 23 m.
The steel tubular pile that forms part of a pile group in a brea- Note that the external diameter d and wall thickness t affect both
sting dolphin (Fig. 7) has been examined probabilistically above ultimate limit state and serviceability limit state functions, by
based on the single performance function of bending failure mode. affecting the moment of inertia I, the Matlock p–y curves (Fig. 7b,
The reliability index is only 1.514. In design, a reliability index of in which B = d = 1.3 m), and the yield moment My = 2ryI/d, where
3.0 or 3.5 is often required. Further, deflection criterion also needs ry is the yield stress (417 MPa) of high-tensile alloy steel.
to be considered. In the following illustrative reliability-based Fig. 7c shows the curves of reliability index b = 3 as a function of
design example, it is assumed that 1.4 m is the maximum tolerable the thickness t and external diameter d of the tubular pile, for the
pile head deflection, which means a secant tilt angle of about 1.4 in deflection performance criterion and pile bending failure mode.
26, or about 3 degrees, with respect to seabed. The configuration at location a (i.e. 32 mm annular wall thickness
Analysis using mean value of PH and mean trend of cu (i.e. and external diameter 1.42 m) yields a reliability index of 3 with
lcu = 150 + 2z, kPa) results in a pile head deflection of 0.986 m respect to both the deflection and the bending moment modes.
(slightly smaller than the 0.994 m of the above mentioned deter- Configurations along line ac will have a reliability index of 3 for
B.K. Low, K.-K. Phoon / Computers and Geotechnics 65 (2015) 30–44 41

bending moment mode, and greater than 3 for deflection mode, of 3.0 against rotational failure about the anchor point ‘‘A’’. The sta-
whereas configurations along line ab means a reliability index of bility formulation was based on the free-earth support method.
3 for deflection mode, and greater than 3 for bending moment Similar to steps (iii) and (iv) in Section 3, the six values under
mode. the n column in Fig. 8 were initially zeros. A trial value of H (e.g.
The next example illustrates the RBD of an anchored sheet pile H = 11 m) was selected such that the performance function value
wall, in which uncertainty in the unit weight is modeled. It will be is positive when n = 0 initially. Excel Solver was then invoked to
shown that the RBD abides by the single-source principle of unit obtain the reliability index b. A few trials of H values, each followed
weight c as in EC7, but is able to determine automatically the with an Excel Solver run, finally obtained the solution in Fig. 8,
design value of the unit weight (namely the value of Characteris- namely that an input total height of 12.15 m would give a reliabil-
tic-value-of- c/Partial-factor-of-c), and thus can play a useful com- ity index b of 3.0 against rotational failure. With this wall height,
plementary role to EC7 which normally specifies 1.0 for the partial the mean-value point is safe against rotational failure, but failure
factor of c, perhaps due to EC7’s uncertainty on whether the design occurs when the mean values descend/ascend to the values indi-
value of unit weight should be higher or lower than its character- cated under the x⁄ column. These x⁄ values denote the design point
istic value when both active pressure and passive resistance on the limit state surface, and represent the most likely combina-
increase with the unit weight of the soil. tion of parametric values that will cause failure. The expected
embedment depth is d = 12.15 – 6.4 – lz = 3.35 m. At the failure
combination of parametric values the design value of z is
8. FORM design of an anchored sheet pile wall z⁄ = 2.9632, and d⁄ = 12.15 – 6.4 – z⁄ = 2.79 m. This corresponds
to an ‘overdig’ allowance of 0.56 m. Unlike Eurocode 7, this ‘over-
The Microsoft Excel constrained optimization approach for reli- dig’ is determined automatically, and reflects uncertainties and
ability-based design of the anchored sheet pile wall in Fig. 8 was sensitivities from case to case in a way that specified ‘overdig’
presented by Low [28], but solved in this study (and obtaining cannot.
the same solution) using the Low and Tang [30] FORM x-via-n pro- The n column indicates that, for the given mean values and
cedure as summarized in Fig. 2 top right. Another example with uncertainties, rotational stability is, not surprisingly, most sensi-
steady seepage condition was presented by Low [27]. Given the tive to /0 and the dredge level (which affects z and d and hence
uncertainties and correlation structure in Fig. 8, one wishes to find the passive resistance). It is interesting to note that at the design
the required total wall height H so as to achieve a reliability index point where the six-dimensional dispersion ellipsoid touches the

Correlation matrix R
x* Mean StDev n u γ γsat qs φ ' δ z
Normal γ 16.204 17 0.85 -0.936 -0.936 γ 1 0.5 0 0.5 0 0
Normal γsat 18.443 20 1 -1.557 -1.258 γsat 0.5 1 0 0.5 0 0

Normal qs 10.276 10 2 0.138 0.138 qs 0 0 1 0 0 0


Normal φ' 33.514 38 2 -2.243 -1.729 φ' 0.5 0.5 0 1 0.8 0
Normal δ 17.293 19 1 -1.707 -0.067 δ 0 0 0 0.8 1 0
Normal z 2.9632 2.4 0.3 1.877 1.877 z 0 0 0 0 0 1

Boxed cells contain Surcharge qs


d* = H − 6.4 − z* equations

Ka Kah Kp Kph H d* PerFn 1.5 m T


0.3 0.249 5.904 5.6372 12.15 2.79 0.00 A
6.4 m
Forces Lever arm Moments β γ, φ', δ
(kN/m) (m) (kN-m/m) 3.00
2
1 -31.1 4.575 -142 Water table
2 -82.8 2.767 -229 1
3 -149 7.775 -1156
z
4 -35.6 8.733 -311

5 189.2 9.72 1839 Dredge level γsat


3

d
5 4

Fig. 8. Design total wall height for a reliability index of 3.0 against rotational failure. Dredge level and hence z and d are random variables.
42 B.K. Low, K.-K. Phoon / Computers and Geotechnics 65 (2015) 30–44

limit state surface, both unit weights c and csat (16.20 and 18.44) tions, and there are parameters in the RBD cases for which
are lower than their corresponding mean values, contrary to the partial factors are not yet specified in the EC7. Presumably the par-
expectation that higher unit weights will increase active pressure tial factors in EC7 already account for some typical values of the
and hence greater instability. This apparent paradox is resolved if coefficient of variation (standard deviation/mean) and probability
one notes that smaller csat will (via smaller c0 ) reduce passive resis- distribution (e.g. Normal or Lognormal distributions) of the actions
tance, smaller /0 will cause greater active pressure and smaller and material properties, and that the resulting EC7 design is meant
passive pressure, and that c, csat, and /0 are logically positively cor- to be sufficiently safe, though not aiming at any specific reliability
related, and thus modeled with positive correlation coefficients in level. With this in mind, one can proceed with the EC7 design using
the correlation matrix of Fig. 8. It is also logical to assume a posi- characteristic values smaller or higher than mean values (for resis-
tive correlation between the angle of friction / of the soil and tance and loading parameters, respectively), and partial factors
the interface friction angle d between the soil and the sheet pile, corresponding to design approaches DA1a, DA1b, DA2 or DA3.
and this also has been modeled in the correlation matrix. The resultant design (for example, foundation width B of Fig. 3,
In this case where some of the six parameters are (logically) or embedment depth of Fig. 8) can then be subjected to FORM anal-
correlated, as shown by the correlation matrix R in Fig. 8, and ysis if statistical information is actually available, to provide some
where uncertainties in unit weights c and csat are modeled and sin- very useful information on the probability of failure and the sensi-
gle-source principle is to be observed, RBD can play a complemen- tivities of the various parameters, which are not conveyed by an
tary role to EC7 in automatically deciding whether the design EC7 Design. One can of course conduct RBD directly if statistical
values of c and csat should be lower or higher than their respective information (standard deviations, mean values, parametric correla-
mean values. tions, probability distributions) is available, to achieve a design for
Clause 2.4.2(9)P of EC7 explains the ‘single-source principle’ in specified target reliability index, which can provide instructive
terms of permanent favorable and unfavorable actions arising from comparisons with EC7 design which does not directly reflect
the same physical source, and suggests that the same partial factor uncertainties and correlation structure.
may be applied to the sum of these actions or to the effect of them.
The RBD example in Fig. 8 applies the same principle logically to a 9. Summary and conclusions
material property, namely that the soil unit weight on both sides of
the wall should be the same, and leaves it to RBD to determine This paper discussed reliability-based design (RBD) for a shal-
whether the design values of the unit weight should be higher or low foundation, a reinforced rock slope, a Norwegian clay slope,
lower than its mean value. Although the design value of csat a laterally loaded pile, and an anchored sheet-pile wall, using the
(18.443 kN/m3) is lower than its mean value of 20 kN/m3 in practical FORM procedure of Low and Tang [30] to explain con-
Fig. 8, for other geometry and embedment depth of the wall and cepts and illustrate the complementary role that FORM can play
for other applications (e.g. the RBD of a Norwegian slope in to the Eurocode 7 design approach. The relatively transparent
Fig. 6) it might be more critical that the design value of the soil unit spreadsheet-automated FORM and SORM procedures can be cou-
weight be higher than its mean value, an issue which RBD can pled with stand-alone numerical packages via bridging techniques
determine automatically from case to case. like response surface methods and neural networks.
The expanding dispersion ellipsoid perspective in the original
8.1. Distinguishing negative from positive reliability indices space of the random variables was presented, as a useful alterna-
tive perspective of the reliability index and the design point.
In Fig. 8, if a trial H value of 10 m is used, and the entire n col- Various insights and interesting features and subtleties of RBD
umn given their initial values of 0.0, the x⁄ column will display the as revealed in the different cases were discussed, testifying to the
respective mean value, and the performance function PerFn exhib- ability of RBD to locate the design point for a target risk level with-
its a negative value of 448.52, meaning that, for the input H of out presuming any partial factors and to reflect parametric sensi-
10 m, the mean value point is already inside the unsafe domain tivities and correlations from case to case automatically. Links
(e.g., to the left of the limit state surface in Fig. 1, but in six-dimen- between RBD and EC7 design were demonstrated, and the comple-
sional space). Upon Solver optimization with constraint PerFn = 0, a mentary role which RBD can play to EC7 were explained in six bul-
b index of 1.34 is obtained, which should be regarded as a negative leted points at the end of Section 4, and further commented in the
index, i.e., 1.34, meaning that the unsafe mean value point is at different RBD cases of subsequent sections. In particular, for differ-
some distance from the nearest safe point on the limit state surface ent mean value and standard deviation of the lateral load in the
that separates the safe and unsafe domains. In other words, the RBD case of footing width of Fig. 3, or for different applications
computed b index can be regarded as positive only if the perfor- (e.g., slope stability or earth retaining walls), the partial factors
mance function (PerFn) value is positive at the mean value point. back-calculated from RBD results could be different. To apply the
For the case in hand, the mean value point (prior to Solver optimi- same specified partial factors of EC7 across different applications
zation) yields a positive PerFn for H > 10.6 m. The computed b or different levels of parametric uncertainty may not imply the
index increases from about 0 (equivalent to a factor of safety equal same target failure probability. In contrast, in an RBD, the same tar-
to 1.0) when H is 10.6 m to 3.0 when H is 12.15 m for the statistical get reliability index (with its implied probability of failure) can be
inputs of Fig. 8. used across different applications and different levels of parametric
Another example for the need to start from a mean-value point uncertainty and correlations, and including asymmetric probabil-
that is in the positive domain before invoking Solver to compute ity distribution if appropriate. Also, if higher reliability is required
reliability index b has been illustrated in step (iii) in connection where the consequence of failure is more significant, a safer design
with the RBD of the foundation width B of Fig. 3. can be obtained by raising the target reliability index. Such flexibil-
ity and automatic incorporation of parametric sensitivities and cor-
8.2. Comparing EC7 design with RBD design, and FORM analysis of EC7 relations into the design process are not found in a Eurocode
design design based on code-recommended partial factors. It is also evi-
dent from the other numerical examples in this study that the
The RBD cases in this study are not directly translatable to EC7 same set of partial factors cannot achieve a uniform reliability
design, because the present form of EC7 does not yet deal with cor- index across different applications or same application with differ-
relation coefficients, spatial variability and probability distribu- ent input parameters (e.g. different coefficient of variation of
B.K. Low, K.-K. Phoon / Computers and Geotechnics 65 (2015) 30–44 43

friction angle). One should hasten to add that EC7 does not sub- [2] Baecher GB, Christian JT. Reliability and statistics in geotechnical
engineering. Chichester, West Sussex, England; Hoboken, NJ: J. Wiley; 2003.
scribe to a reliability-based design philosophy and hence there is
[3] Bowles JE. Foundation analysis and design. 5th ed. McGraw-Hill; 1996.
no intent to produce a uniform reliability index across different [4] Chan CL, Low BK. Practical second-order reliability analysis applied to
scenarios. Nonetheless, coefficients of variation and cross/auto- foundation engineering. Int J Numer Anal Meth Geomech
correlations are statistical features of measured data and one 2012;36(11):1387–409.
[5] Chan CL, Low BK. Probabilistic analysis of laterally loaded piles using response
may conclude that EC7 is not sensitive to the statistical aspects surface and neural network approaches. Comput Geotech 2012;43:101–10.
of measured data. When statistical data are available, reliability- [6] Ching JY, Phoon KK, Hu YG. Efficient evaluation of reliability for slopes with
based design can be compared to EC7 design to illuminate the circular slip surfaces using importance sampling. J Geotech Geoenviron Eng,
ASCE 2009;135(6):768–77.
impact of additional data on design. Finally, even though uncer- [7] Ching JY, Phoon KK, Hu YG. Observations on limit equilibrium based slope
tainties are implicitly considered in EC7, it is worth pointing out reliability problems with inclined weak seams. J Eng Mech, ASCE
that the practice of assigning a single numerical value to each par- 2010;136(10):1220–33.
[8] Ching JY, Phoon KK. Modeling parameters of structured clays as a multivariate
tial factor implies that only a narrow range of coefficients of vari- normal distribution. Can Geotech J 2012;49(5):522–45.
ation has been considered. Soil property evaluation [9] Ching JY, Phoon KK. Multivariate distribution for undrained shear strengths
methodologies are more diverse (to suit diverse site conditions under various test procedures. Can Geotech J 2013;50(9):907–23.
[10] Ching JY, Phoon KK, Chen CH. Modeling piezocone cone penetration (CPTU)
and practices evolved to suit these conditions) in contrast to struc- parameters of clays as a multivariate normal distribution. Can Geotech J
tural property evaluation methodologies. For situations involving 2014;51(1):77–91.
coefficients of variation higher than those implied in EC7, it may [11] Ching JY, Phoon KK. Transformations and correlations among some clay
parameters – the global database. Can Geotech J 2014;51(6):663–85.
be judicious to apply RBD as an additional design check.
[12] Ching JY, Phoon KK. Correlations among some clay parameters – the
Reliability-based design/analysis involving spatially autocorre- multivariate distribution. Can Geotech J 2014;51(6):686–704.
lated soil properties was illustrated and discussed for a Norwegian [13] Der Kiureghian A, Liu PL. Structural reliability under incomplete probability
slope (Section 6) and for a laterally loaded pile (Section 7). In the information. J Eng Mech ASCE 1986;112(1):85–104.
[14] DerKiureghian A, Lin HZ, Hwang SJ. Second-order reliability approximations. J
latter case, RBD with respect to serviceability limit state was also Eng Mech ASCE 1987;113(8):1208–25.
implemented. [15] Ditlevsen O. Uncertainty modeling: with applications to multidimensional
In the RBD of the embedment depth of an anchored sheet pile civil engineering systems. New York: McGraw-Hill; 1981.
[16] EN 1997-1, Eurocode 7: geotechnical design, Part 1: general rules.
wall in Section 8, the soil unit weight above water table affects only [17] Frank R, Bauduin C, Driscoll R, Kavvadas M, Krebs Ovesen N, Orr T, et al.
active pressure and thus is of the nature of a loading parameter, Designers’ guide to EN 1997-1 Eurocode 7: geotechnical design – general
whereas the unit weight csat below water table affects (via c0 = csat - rules. Thomas Telford; 2005.
[18] Haldar A, Mahadevan S. Probability, reliability and statistical methods in
 cw) both the active pressure behind the wall and the passive engineering design. New York: John Wiley; 1999.
resistance below the dredge level and is hence of the nature of both [19] Hasofer AM, Lind NC. Exact and invariant second-moment code format. J Eng
a load and a resistance parameter. The complexly intertwined Mech 1974;100:111–21. ASCE, New York.
[20] Hoek E. Practical rock engineering. <http://www.rocscience.com/education/
parametric sensitivities and correlations make it difficult to apply hoeks_corner>; 2007.
partial factors consistently. In a reliability-based design (such as [21] Li D-Q, Wu S-B, Zhou C-B, Phoon KK. Performance of translation approach for
the case in Fig. 8) one does not prescribe the ratios mean/design modeling correlated non-normal variables. Struct Saf 2012;39:52–61.
[22] Low BK, Tang Wilson H. Efficient reliability evaluation using spreadsheet. J Eng
value—such ratios, or ratios of (characteristic values)/design value,
Mech, ASCE 1997;123(7):749–52.
are prescribed as partial factors in EC7 limit state design—but [23] Low BK, Tang WH. Reliability analysis of reinforced embankments on soft
leaves it to the expanding dispersion ellipsoid to seek the most ground. Can Geotech J 1997;34(5):672–85.
probable failure point on the limit state surface (similar to Fig. 1 [24] Low BK, Teh CI, Tang Wilson H. Stochastic nonlinear p–y analysis of laterally
loaded piles. In: Proceedings of the eight international conference on
but in multi-dimensional random variable space), a process which structural safety and reliability. Newport Beach, California: A.A.Balkema
automatically reflects the complexly intertwined parametric sensi- Publishers; 2001. 17–22 June 2001, 8p.
tivities and correlations. Besides, one can associate a probability of [25] Low BK. Practical probabilistic slope stability analysis. In: Proc, soil and rock
America. Cambridge, Massachusetts: M.I.T., Verlag Glückauf GmbH Essen, vol.
failure for each target reliability index value. The ability to seek the 2. 2003. p. 2777–84. (http://www.ntu.edu.sg/home/cbklow/).
most-probable design point without presuming any partial factors [26] Low BK, Tang Wilson H. Reliability analysis using object-oriented constrained
and to reflect sensitivities from case to case automatically is a optimization. Structural Safety, vol. 26, No. 1. Amsterdam: Elsevier Science
Ltd.; 2004. p. 69–89.
desirable feature of the reliability-based design approach. [27] Low BK. Reliability-based design applied to retaining walls. Geotechnique
It was shown that for some geotechnical problems, performing 2005;55(1):63–75.
SORM after FORM can improve the accuracy of the estimated fail- [28] Low BK. Practical reliability approach in geotechnical engineering. In:
Proceedings of geocongress 2006: geotechnical engineering in the
ure probability, but a pragmatic approach on the adequacy of the information technology age, Atlanta, Georgia, USA, 26 February – 1 March
approximate failure probability estimated from FORM b index 2006, 6 pages in CD-ROM, ASCE; 2006.
was also suggested and discussed in the context of the results of [29] Low BK. Reliability analysis of rock slopes involving correlated nonnormals. Int
J Rock Mech Min Sci 2007;44(6):922–35.
a reinforced rock slope in Section 5. This pragmatic approach does
[30] Low BK, Tang Wilson H. Efficient spreadsheet algorithm for first-order
not require performing SORM, and may thus reduce the hurdles reliability method. J Eng Mech, ASCE 2007;133(12):1378–87.
confronting practitioners in using reliability-based design. The [31] Low BK, Lacasse S, Nadim F. Slope reliability analysis accounting for spatial
more rigorous approach involving both FORM and SORM can be variation. Georisk: assessment and management of risk for engineered
systems and geohazards, vol. 1, No. 4. London,: Taylor & Francis; 2007. p.
used, if desired, after some familiarity has been gained with the 177–89.
pragmatic FORM-only approach. [32] Low BK. Practical reliability approach using spreadsheet. In: Phoon KK, editor.
The spreadsheet-automated [30] FORM procedure has been Reliability-based design in geotechnical engineering-computations and
applications. Taylor and Francis; 2008. p. 134–68 [Chapter 3].
used in this study for the purpose of explaining FORM and the intu- [33] Low BK, Zhang J, Tang Wilson H. Efficient system reliability analysis illustrated
itive dispersion ellipsoidal perspective and the complementary for a retaining wall and a soil slope. Comput Geotech 2011;38(2):196–204.
role of FORM reliability-based design to EC7 design approach. In Elsevier.
[34] Low BK, Einstein HH. Reliability analysis of roof wedges and rockbolt forces in
practice RBD can be done using either the [30] FORM procedure tunnels. Tunn Undergr Space Technol 2013;38:1–10.
or other available FORM packages. [35] Lü Q, Low BK. Probabilistic analysis of underground rock excavations using
response surface method and SORM. Comput Geotech 2011;38:1008–21.
Elsevier.
References [36] Madsen HO, Krenk S, Lind NC. Methods of structural safety. Englewood Cliffs,
NJ: Prentice-Hall; 1986.
[1] Ang HS, Tang WH. Probability concepts in engineering planning and design. [37] Matlock H. Correlations for design of laterally loaded piles in soft clay. In: Proc,
Decision, risk, and reliability, vol. 2. New York: John Wiley; 1984. offshore technology conference, Houston, Texas, 1970, Paper OTC 1204; 1970.
44 B.K. Low, K.-K. Phoon / Computers and Geotechnics 65 (2015) 30–44

[38] Melchers RE. Structural reliability analysis and prediction. 2nd ed. New [43] Tomlinson MJ. Pile design and construction practice. 4th ed. London: E & FN
York: John Wiley; 1999. Spon; 1994.
[39] Rackwitz R, Fiessler B. Structural reliability under combined random load [44] Tomlinson MJ. Foundation design and construction. 6th ed. Longman
sequences. Comput Struct 1978;9(5):484–94. Scientific; 1995.
[40] Shinozuka M. Basic analysis of structural safety. J Struct Eng ASCE [45] Veneziano D. Contributions to second moment reliability. In: Research Report
1983;109(3):721–40. No. R74–33. Department of Civil Engineering, Cambridge, Massachusetts: MIT;
[41] Shooman ML. Probabilistic reliability, an engineering approach. New York, 1974.
N.Y.: McGraw-Hill Book Co.; 1968. [46] Whitman RV. Evaluating calculated risk in geotechnical engineering. J Geotech
[42] Simpson B, Driscoll R. Eurocode 7 – a commentary. Watford: Construction Eng ASCE 1984;110(2):145–88.
Research Communications; 1998.

Você também pode gostar