Robust Geotechnical Design of Drilled Shafts in Clay Using …hsein/wp-content/uploads/2017/... ·...
Transcript of Robust Geotechnical Design of Drilled Shafts in Clay Using …hsein/wp-content/uploads/2017/... ·...
1
Robust Geotechnical Design of Drilled Shafts in Clay Using Spreadsheet 1
2
Wenping Gong, S.M.ASCE1, Sara Khoshnevisan
2, C. Hsein Juang, F.ASCE
3, 3
Sez Atamturktur, M.ASCE3 4
5
Abstract 6
This paper presents an efficient robust geotechnical design (RGD) approach that 7
considers performance requirements, design robustness, and cost efficiency simultaneously. 8
In this paper, the design robustness is measured with the variation of the performance function 9
of concern that is established for a reliability analysis. In the proposed RGD approach, the 10
performance requirement of the system is evaluated using reliability analysis. Thus, the 11
evaluation of the design robustness and the evaluation of the performance requirement share 12
common computational steps, referred to herein as computational “coupling.” This coupling 13
for computational efficiency is a significant feature of the proposed RGD approach. Within 14
the framework of the proposed RGD, the design robustness, cost efficiency, and performance 15
requirements are considered simultaneously by means of multi-objective optimization. 16
Furthermore, in this paper a practical procedure is developed for such optimization using 17
Solver, an optimization feature of the Excel spreadsheet software. Through an example of the 18
design of a drilled shaft in clay, the effectiveness and significance of this new RGD approach 19
is demonstrated. The results show that the hard-to-control variability resulting from 20
construction variation, loading conditions, model errors, and geotechnical parameters in the 21
design of drilled shafts in clay can be effectively considered with the proposed RGD approach. 22
23 Keywords: Drilled shaft; optimization; Pareto front; performance function; reliability analysis; 24
robust geotechnical design. 25
2
26
_______________ 27
1Research Assistant, Glenn Department of Civil Engineering, Clemson University, Clemson, 28
SC 29634; and Department of Geotechnical Engineering, Tongji University, Shanghai 200092, 29
China. 30
31 2Research Assistant, Glenn Department of Civil Engineering, Clemson University, Clemson, 32
SC 29634; and Department of Geotechnical Engineering, Tongji University, Shanghai 200092, 33
China. 34
35 3Glenn Professor, Glenn Department of Civil Engineering, Clemson University, Clemson, SC 36
29634. 37
38 4Associate Professor, Glenn Department of Civil Engineering, Clemson University, Clemson, 39
SC 29634. 40
41
42
3
Introduction 43
Performance or response of a geotechnical system under loading can be either over- or 44
under-estimated due to the uncertainty in soil parameters, loading conditions, analysis model 45
errors, and construction variation. To cope with the uncertainty in the predicted system 46
response, the engineer often selects a conservative design that may be cost inefficient. While 47
the reliability-based methods are capable of considering explicitly the uncertainty in input 48
parameters and solution models (e.g., Wu et al. 1989; Christian et al. 1994; Baecher and 49
Christian 2003; Phoon et al. 2003; Fenton and Griffiths 2008), reliability analysis and design 50
requires an accurate statistical characterization of such uncertainty, which is often challenging 51
in practice for various reasons. The results of a reliability-based design (RBD) might be 52
questionable if an accurate statistical characterization of the uncertainty cannot be obtained, 53
e.g., the under-estimated variability of soil parameters can lead to an under-design while the 54
over-estimated variability of soil parameters can lead to an over-design (Juang et al. 2013a). 55
To deal with the uncertainty in the estimated statistics of soil parameters (called noise 56
factors herein), Juang et al. (2013a&b&c) proposed a reliability-based robust geotechnical 57
design (RGD) approach. The goal of RGD is to reduce the effect of the uncertainty in the 58
noise factors on the predicted performance of the geotechnical system. Robust design is 59
achieved through adjusting the easy-to-control design parameters (i.e., geometry parameters) 60
so that the variation of the system performance or response is reduced to an extent that the 61
designed system is considered insensitive to, or robust against, the variability in the hard-to-62
control noise factors (i.e., estimated statistics of soil parameters). Here, robust is interpreted 63
as a kind of capability that measures the insensitivity of the system response to the variability 64
4
in the noise factors; accordingly, robustness is referred to as an ability of the system to resist 65
the inevitable and unforeseen variation of noise factors. 66
RGD is not a new concept in engineering; rather, it is adapted from the well-67
established robust design methodology, initiated by Taguchi (1986) in the field of quality 68
engineering. In fact, numerous applications of this concept have been reported in various 69
engineering fields (e.g., Tsui 1992; Chen et al. 1996; Doltsinis et al. 2005; Zhang et al. 2005; 70
Park et al. 2006; Beyer and Sendhoff 2007; Marano et al. 2008; Lee et al. 2010). Adapting the 71
concept of robust design to the field of geotechnical engineering, however, requires some 72
refinements as the level of uncertainty in input parameters, solution model, and construction 73
variation is generally much higher than that in other fields of engineering. Therefore, how to 74
select the most effective and efficient approach is a critical issue that must be resolved for a 75
successful implementation of robust design principles in geotechnical engineering. 76
In the previous study, Juang et al. (2013a) presented a reliability-based RGD approach, 77
in which the performance of drilled shafts is analyzed using reliability-based methods, and the 78
variation of the system performance, in terms of failure probability, is utilized as a measure of 79
the design robustness. The goal of RGD is to optimize the design robustness and the cost 80
simultaneously, while satisfying the safety or performance requirements. To this end, a multi-81
objective optimization is often implemented in the RGD. The reliability-based RGD approach 82
by Juang et al. (2013a) involves three loops, one for computing failure probability using 83
reliability methods, one for computing the variation of the failure probability considering the 84
uncertainty in the estimated statistics of input parameters, and the last one for conducting 85
multi-objective optimization. Although this reliability-based RGD approach is shown as an 86
effective design tool in geotechnical engineering (i.e., Juang et al. 2013a; Juang and Wang 87
5
2013; Wang et al. 2013), its multi-objective optimization-based implementation is quite 88
computationally demanding. A more efficient RGD approach is desirable. 89
The efficiency of a robust design is affected by how the system response is measured 90
and computed, and how the design robustness is defined. For example, Juang et al. (2013c) 91
described a RGD approach with the deformation, in lieu of failure probability, as the system 92
response so that the evaluation of the variation of the system response involves only one loop 93
of iterative procedure rather than two, which yields significant savings in computation time. 94
Indeed, the type of chosen system response also dictates the choice and efficiency of a 95
particular robustness measure. In literature, the commonly adopted robust design approaches 96
include Taguchi’s method (e.g., Taguchi 1986; Phadke 1989), worst-case approach (e.g., 97
Nagy and Braatz 2004; Zhu and Fukushima 2009), mean-variation compromising approach 98
(e.g., Chen et al. 1996; Park et al. 2006), and feasibility robustness approach (e.g., Parkinson 99
1993; Du and Chen 2000). Interested readers are referred to the literature for detailed 100
information, and to Juang et al. (2013b), Wang et al. (2013), and Gong et al. (2014) for 101
example applications in geotechnical engineering. 102
In this paper, the RGD approach proposed by Juang et al. (2013a) is further refined. 103
The main objective is to enhance the RGD approach as an effective and efficient design tool. 104
Several new features are introduced into the refined RGD framework. First, a fundamentally 105
sound and intuitive measure of the design robustness based on the variation of the 106
performance function is presented. Second, greater computation efficiency is achieved by 107
“coupling” the evaluation of design robustness with the evaluation of the performance 108
requirement using reliability methods (note: the word “coupling” here means that the two 109
evaluation procedures share common computational steps). Third, an effective procedure to 110
6
transform the multi-objective optimization into a series of single-objective optimizations is 111
developed so that the entire RGD process can be easily implemented in the user friendly 112
Excel spreadsheet environment. The improved RGD approach presented in this paper is not 113
only efficient but also user friendly and practical. 114
This paper is organized as follows. First, the formulation of a robustness measure, in 115
the form of the variation of the system performance function, is presented. Next, a simplified 116
reliability-based approach to couple the evaluation of design robustness with the evaluation of 117
the performance requirement is discussed. Then, the framework of the proposed efficient 118
RGD approach is formulated. Finally, an application of the proposed RGD approach to the 119
design of a drilled shaft in clay is demonstrated, where all formulations are implemented in 120
the Excel spreadsheet environment. 121
122
Variation of the Performance Function as a Robustness Measure 123
Robust design is aimed at deriving an optimal design, represented by a set of design 124
parameters (d), the system response of which is insensitive to, or robust against, the variation 125
of noise factors (θ) (e.g., Taguchi 1986; Tsui 1992; Chen et al. 1996; Beyer and Sendhoff 126
2007; Juang et al. 2013a&b&c; Gong et al. 2014). In principle, the design robustness can be 127
raised by reducing the variation of the system response that arises from the variation of noise 128
factors. In this paper, the variation of the response or performance of a geotechnical system, 129
which may be obtained through a reliability analysis, is used as a robustness measure. Use of 130
such robustness measure enables a “coupling” of the evaluation of design robustness with the 131
evaluation of the system performance. 132
Formulation of the performance function in the context of robust design 133
7
Within the context of robust design, the input parameters are classified into two 134
categories, one is the easy-to-control design parameters (d) and the other is the hard-to-135
control noise factors (θ). For a geotechnical system with design parameters of d and noise 136
factors of θ as inputs, the performance function, denoted as g(d, θ), can be set up as: 137
( , ) ( , ) ( , )g R S d d d (1) 138
where R(d, θ) and S(d, θ) are the resistance term and load term, respectively. The performance 139
of a design (d) is considered unsatisfactory if g(d, θ) < 0. Noise factors such as uncertain soil 140
parameters, loading conditions, model errors, and construction variation are treated as random 141
variables in a reliability analysis. To meet the performance requirement, the evaluated failure 142
probability (Pf) of a design, referred to herein as the probability of unsatisfactory performance, 143
must be less than a specified target value (PT). 144
Variation of the performance function as a robustness measure 145
With the performance function g(d, θ) defined in Eq. (1), the design robustness can be 146
intuitively measured by its variation (i.e., standard deviation). Fundamentally, a greater 147
variation of the performance function indicates a higher level of sensitivity of the system 148
response to the variation of noise factors. Using Taylor series expansions, the design 149
robustness of a design (d), in the form of the standard deviation of the performance function, 150
can be approximated as (e.g., Ang and Tang 2004; Zhang et al. 2011): 151
[ ] Tg CG G (2) 152
where Cθ is the covariance matrix among noise factors (θ), and G is the gradient vector of the 153
performance function to noise factors (θ), formulated as: 154
8
1 2
( , ) ( , ) ( , ), , ,
n
g g g
μ μ μ
d d dG (3) 155
where θ presents the mean of noise factors (θ). Accordingly, a lower standard deviation of 156
the performance function ([g]) signals a higher degree of design robustness. 157
Two possible approaches to improve design robustness 158
The measure of design robustness expressed in Eq. (2) is affected by two terms, the 159
gradient vector (G) and the covariance (Cθ). Hence, two approaches are possible to improve 160
the design robustness, either by minimizing the gradient (referred to as Approach 1 in Figure 161
1) or by minimizing the variation of noise factors (referred to as Approach 2 in Figure 1). 162
The first approach is to tune the design parameters (d) such that the gradient (G) is 163
reduced for a given variation of noise factors (). This is precisely the approach taken in the 164
current robust design (e.g., Taguchi 1986; Chen et al. 1996; Nagy and Braatz 2004; Juang et 165
al. 2013a&b&c; Liu et al. 2013; Gong et al. 2014). The second approach seeks to raise the 166
design robustness by reducing the variation of noise factors (θ). 167
168
MFOSM-Based Procedure for Reliability Analysis of System Performance 169
Given noise factors, the question of whether the performance requirement of a system 170
is satisfied or not is difficult to evaluate with certainty using deterministic methods. Thus, it is 171
appropriate to evaluate the system performance using probabilistic approaches. Although such 172
evaluation can be done using rigorous methods such as Monte Carlo simulation (MCS) and 173
first order reliability method (FORM), the computational demand of the RGD optimization 174
that involves MCS or FORM is often prohibitive. Therefore, the mean first order second 175
9
moment method (MFOSM), a much simpler reliability method, is employed herein to 176
evaluate the system performance within the proposed RGD framework. 177
Mean first order second moment method (MFOSM) 178
Within the context of MFOSM, the mean (E[g]) of the performance function g(d, θ) 179
can be approximated as follows (e.g., Ang and Tang 2004; Zhang et al. 2011): 180
[ ] ( , )E g g μ d (4) 181
While the standard deviation ([g]) of the performance function g(d, θ) is readily available 182
from the formulation of design robustness (see Eq. 2), the reliability index () of the system 183
performance, g(d, θ) > 0, can be computed as follows, with an assumption that g(d, θ) follows 184
a normal distribution (e.g., Zhao and Ono 2001; Ang and Tang 2004; Zhang et al. 2011): 185
[ ]
[ ]
E g
g
(5) 186
If desired, the failure probability (Pf) with respect to the system performance 187
requirement, or the probability of unsatisfactory performance, can be computed according to 188
the obtained reliability index (), which is expressed as (e.g., Zhao and Ono 2001; Ang and 189
Tang 2004; Zhang et al. 2011): 190
( )fP (6) 191
It is noted that through the common formulation of the variation of the performance 192
function, the evaluation of the design robustness and the evaluation of the system 193
performance, two main procedures in the RGD, share common computational steps or are 194
“coupled” from the perspective of computational effort. This “coupling” for computational 195
efficiency is a significant feature of the efficient RGD approach proposed herein. 196
Evaluation of system performance using MFOSM-based procedure 197
10
Though MFOSM is often favored for its simplicity over more rigorous methods such 198
as MCS and FORM, the accuracy of MFOSM might be questionable in many cases. To 199
overcome the dilemma between accuracy and computation efficiency, the procedure presented 200
by Zhang et al. (2011) is employed here. Zhang et al. (2011) pointed out that the reliability 201
index (M) that is computed using MFOSM is usually highly correlated with the reliability 202
index (T) that is computed using more sophisticate and accurate reliability methods such as 203
MCS and FORM. Thus, a mapping function can be established for a given problem: 204
M T( )f (7) 205
Note that the reliability index obtained using MCS or FORM is considered a close 206
approximation of the “true” reliability index, and thus the superscript “T” (for “true”) is used 207
for the reliability index computed with MCS or FORM. Similarly, the superscript “M” is for 208
the reliability index computed using MFOSM. The mapping function defined in Eq. (7) can 209
be constructed through the following procedures. First, compute the values of M using 210
MFOSM and T using MCS or FORM for a series of designs in a given design space of a 211
given problem. Second, carry out the least-square regression analysis to develop a mapping 212
function of M = f(T
). Third, validate the obtained mapping function. 213
Once the mapping function shown in Eq. (7) is successfully established in a design 214
space of a given problem, the target reliability index (M
T ) for system performance 215
requirement, within the context of MFOSM analysis, can be set up as M
T = f(T), where T is 216
the specified target reliability index (Zhang et al. 2011). Thus, the MFOSM analysis along 217
with the calculated M
T can readily be used to analyze the system performance requirement 218
without the need of performing MCS or FORM analysis. 219
11
Framework for Efficient Robust Geotechnical Design 220
Figure 2(a) illustrates the robust design implemented as a multi-objective optimization 221
problem that considers design robustness and cost as objectives while the system performance 222
requirement and the design space are treated as constraints. As noted, the robust design seeks 223
an optimal design in the design space that the design robustness is maximized and the cost (C) 224
is minimized simultaneously, while the system performance requirements are satisfied. 225
A simplified optimization procedure 226
Typically, the design robustness and the cost efficiency are two conflicting objectives. 227
Thus, a single best solution that is optimal with respect to both objectives simultaneously is 228
not attainable; rather, a set of non-dominated optimal solutions often exist that are superior to 229
all others in the design space, but within which, none of them are superior or inferior to others. 230
These non-dominated solutions form a Pareto front. As can be seen later, all these non-231
dominated optimal designs on the Pareto front are equally optimal in the sense that no 232
improvement can be achieved in one objective without worsening in the other objective. 233
Figure 3 plots a conceptual illustration of the non-dominated solution and Pareto front in a 234
robust design optimization. The multi-objective optimization problem shown in Figure 2(a) 235
may be solved using genetic algorithms such as Non-dominated Sorting Genetic Algorithm 236
version II, NSGA-II (Deb et al. 2002). Although NSGA-II is a fast algorithm, such multi-237
objective optimization requires the knowledge of genetic algorithms and considerable 238
programming skills, and is usually computationally demanding since it also requires 239
reliability analysis within the framework of multi-objective optimization. 240
In this paper, the multi-objective optimization problem shown in Figure 2(a) is solved 241
through a series of single-objective optimizations, as shown in Figure 2(b). Unlike the 242
12
weighted-sum-of-objectives approach (e.g., Kim and De Weck 2005; Marler and Arora 2010), 243
the cost, in the proposed approach, is modeled as a constraint instead of an objective (along 244
with the constraints of design space and system performance requirements). Using the single-245
objective optimization setting in Figure 2(b), the most robust design within a given acceptable 246
cost level (CTj) can readily be screened out using an Excel spreadsheet. As depicted in Figure 247
3, the resulting optimal design, denoted as (Cj, j[g]), would be a non-dominated solution on 248
the Pareto front that would have been obtained using multi-objective optimization in Figure 249
2(a). Through this approach, the Pareto front can be collectively formed by a series of optimal 250
designs obtained from single-objective optimizations. 251
The approach of transforming multi-objective optimization into a series of single-252
objective optimizations is inspired by previous studies (e.g., Haimes et al. 1971; Hämäläinen 253
and Mäntysaari 2002). While algorithms such as NSGA-II are generally more suitable for 254
multi-objective optimization, it is equally efficient to implement it through a series of single-255
objective optimizations when the optimization involves only two objectives (in our case, cost 256
and robustness). Since the geotechnical problems such as the design of drilled shafts can 257
generally be modeled as a bi-objective optimization problem, the proposed approach is 258
equally efficient. The proposed approach, however, has a tremendous advantage as it can be 259
easily implemented in an Excel spreadsheet without the knowledge of genetic algorithms and 260
considerable programming skills. As will be shown later, the proposed approach is efficient 261
and able to yield practically the same results as those obtained with NSGA-II multi-objective 262
optimization algorithm. Thus, the proposed RGD approach has a potential as a practical 263
design tool for engineering applications. 264
Procedures to implement the RGD approach 265
13
The procedures to implement the proposed efficient robust geotechnical design (RGD) 266
approach are summarized in the following four main steps: 267
Step 1: Characterize the problem of concern. Here, the system performance of concern, 268
noise factors, and design parameters are identified; meanwhile, the design space, requirements 269
of system performance, design robustness, and cost are formulated. 270
Step 2: Construct the mapping function that relates T (computed using MCS or 271
FORM) to M (computed using MFOSM) for the domain problem characterized in Step 1. 272
This step enables an accurate evaluation of the system performance using MFOSM analysis. 273
Step 3: Carry out the robust design optimization considering the design robustness, 274
cost efficiency, and design constraints using the optimization procedure shown in Figure 2(b). 275
Here, the most robust design within a given cost level is identified using Excel “Solver”. 276
Step 4: Construct the Pareto front to reveal the tradeoff between design robustness and 277
cost efficiency in the given design space. The Pareto front may be constructed through 278
following steps. First, determine the least cost design and the most robust design among all 279
feasible designs (i.e., those that satisfy the performance requirements) in the design space and 280
denote the costs for these designs as CL and CR, respectively. Second, divide the resulting cost 281
interval of [CL, CR] into a series of cost levels, denoted as CT = {CT1, CT2, CT3, …, CTn}. Third, 282
determine the most robust design within each cost level. Fourth and lastly, plot all the 283
resulting designs, in terms of design robustness versus cost, which yields a Pareto front. 284
285
Illustrative Example Robust Design of A Drilled Shaft in Clay 286
To demonstrate the proposed efficient robust geotechnical design (RGD) approach, a 287
design example published by ETC10 (2009) for the design of a drilled shaft in clay, shown in 288
14
Figure 4, is adopted here. The RGD optimization of this example drilled shaft is accomplished 289
with the following steps. 290
Step 1: Model characterization and parameters setting 291
Two fundamental requirements of system performance in the design of drilled shafts 292
in clay are bearing capacity (safety requirement) and settlement (serviceability requirement). 293
These performance requirements are generally specified in terms of ultimate limit state (ULS) 294
and serviceability limit state (SLS), respectively. The performance functions with respect to 295
ULS and SLS, in terms of g1(d, θ) and g2(d, θ), are formulated respectively as follows: 296
1 ULS d l( , )g Q F F d (8) 297
2 SLS d l( , )g Q F F d (9) 298
where QULS represents the ULS bearing capacity; QSLS is the SLS bearing capacity that is 299
derived based on a specified maximum allowable settlement of sa (e.g., taken as 20 mm in this 300
example); and, Fd and Fl represent the dead load and live load on the drilled shaft, 301
respectively. For drilled shafts in clay, the ULS bearing capacity, QULS, can be assessed using 302
the formulations proposed by Phoon and Kulhawy (2005): 303
ULS s bQ Q Q W (10) 304
where Qs and Qb are the side resistance and tip resistance, respectively, and W is the weight of 305
drilled shaft. The side resistance, Qs, in Eq. (10) is estimated as (Phoon and Kulhawy 2005): 306
s u1 u10.33 0.17 ( p ) πaQ c BDc (11) 307
where cu1 is the average undrained shear strength over the shaft length; pa is the atmosphere 308
pressure (taken as 100 kPa); B and D are the diameter and length of drilled shaft, respectively; 309
15
and, is a normal random variable with a mean of 0 and a standard deviation of 0.12. The tip 310
resistance, Qb, in Eq. (10) is computed as (Phoon and Kulhawy 2005): 311
1.12
2
b tc u2 cd cr6.17 π 4Q M c q B
(12) 312
where cu2 is the average undrained shear strength within a depth of B below the tip; cd and 313
cr are modifiers for the shaft depth and soil rigidity, respectively; q is the total vertical stress 314
at shaft dip, taken as D ( is the total unit weight of clay); and, Mtc is a lognormal random 315
variable with a mean of 0.45 and a standard deviation of 0.13. The depth modifier in Eq. (12) 316
is evaluated as (Phoon and Kulhawy 2005): 317
1
cd 1 0.33tan D B (13) 318
The rigidity modifier in Eq. (12), cr , is generally taken as 1.0. Finally, the SLS bearing 319
capacity of drilled shafts in clay, QSLS, can be estimated as (Dithinde et al. 2010): 320
SLS a
ULS a
Q s
Q a bs
(14) 321
where a and b are two hyperbolic curve-fitting parameters for the normalized load-settlement 322
curve, which can be statistically represented with lognormal variables; and, sa is the maximum 323
allowable settlement of drilled shafts. The mean of a and b are 2.79 mm and 0.82, 324
respectively, while the standard deviations are 2.04 mm and 0.09, respectively; and the 325
correlation between these two fitting parameters is -0.801. It is noted that the model errors 326
with respect to both ULS bearing capacity and SLS bearing capacity are explicitly considered 327
through the model parameters of , Mtc, a, and b. While the aforementioned empirical models 328
are employed here to estimate the ULS and SLS bearing capability of drilled shafts in clay, it 329
is not the limitation of the proposed RGD approach. Indeed, the more sophisticated finite 330
16
element method (FEM) that considers the non-linear soil-structure interaction and non-linear 331
and time-dependent behavior of soil can also be used. In which, the response surface method 332
(e.g., Faravelli 1989; Li et al. 2011) might be adopted to predict the ULS and SLS bearing 333
capability of drilled shafts in clay and to calibrate the associated model uncertainty. 334
For this drilled shaft example, a well-documented and well-studied example, the water 335
table is at the ground surface and the unit weight of clay () is 20 kN/m3. Hence, the 336
uncertainty of geotechnical parameters mainly comes from the undrained shear strength (cu). 337
In reference to Figure 5, the undrained shear strength of clay, obtained from undrained triaxial 338
tests, increases linearly with depth; the normalized undrained shear strength, cn = cu / z, is 339
sufficiently represented by a lognormal variable with a mean of 13.04 and a standard 340
deviation of 4.15. The adopted lognormal distribution of cn is validated, with 95% confidence 341
level, using Chi-Square test (2) and the statistics are estimated using the maximum likelihood 342
principle. The unit weight of concrete, an essential input parameter for evaluating QULS and 343
QSLS, is treated as a constant at 24 kN/m3. To account for the uncertainty in loading conditions, 344
both the dead load (Fd) and live load (Fl) are represented as lognormal variables. Given the 345
characteristic value of load F (from Eurocode 7: Fdc = 300 kN and Flc = 150 kN), the mean of 346
load F, denoted as (E[F]), can be approximated as follows (Wang et al. 2011): 347
c
F
[ ]1 1.645
FE F
(15) 348
where Fc represents the characteristic value of load F, and F is the coefficient of variation of 349
load F. With an assumption that the coefficients of variation of the dead load (Fd) and live 350
load (Fl) are 0.1 and 0.18, respectively (Zhang et al. 2011), the mean of the dead load (Fd) and 351
live load (Fl) are calculated as 258 kN and 116 kN, respectively. 352
17
Due to construction imperfections and unforeseen geological conditions, differences 353
between the designed geometry and the constructed geometry of a drilled shaft should be 354
expected. The effect of such discrepancy on the pile capacity can be significant (O’Neill 2001; 355
Iskander et al. 2003; Poulos 2005). Yet, a quantitative index measuring such differences is 356
rarely reported in literature. In the context of robust design, the as-built geometry parameters 357
of a drilled shaft (i.e., BT and DT) are herein modeled as lognormal variables. The mean of 358
these parameters (i.e., BT and DT) are assumed to be the same as the parameters of the 359
designed geometry (i.e., B and D), while the standard deviations of these parameters are both 360
set to 0.05 m; further, the correlation coefficient between the two geometry parameters is 361
assumed as T T, 0.5B D . The standard deviations and the correlation coefficient assumed 362
above are just an example for the purpose of illustrating the effect of construction noise. In 363
practice, other distribution type and different statistics can be adopted based on local 364
experience to yield a more preferred design. Additional discussion on the effect of the level of 365
construction variation is presented later. 366
In summary, the geotechnical parameters, loading parameters, construction variations, 367
and model errors are explicitly included in the robust design with noise factors of cn, Fd, Fl, 368
BT, DT, , Mtc, a, and b. The designed geometry parameters (i.e., B and D) are treated as 369
design parameters. For illustration purpose, the design space is set up as a continuous space, 370
DS = {(B, D) | B [0.3 m, 0.6 m], D [10.0 m, 25.0 m]}. Within the framework of robust 371
design, the target reliability indexes (i.e., T1 and T2) with respect to ULS performance and 372
SLS performance are set as 3.2 and 2.6, respectively, to ensure an adequate system 373
performance. These target reliability indexes correspond to the target failure probabilities (i.e., 374
PT1 and PT2) of 6.910-4
and 4.710-3
, respectively. For this design example, the study 375
18
performed by Orr et al. (2011) found that the critical factor that controls the design is the ULS 376
performance requirement. Thus, the variation of the performance function with respect to 377
ULS performance, [g1(d, )], is adopted herein to measure the design robustness. In this 378
paper, the cost of the drilled shaft (C) is simply represented with the volume of concrete 379
(B2D/4), although other, more sophisticated cost estimate models may be adopted. Figure 6 380
illustrates all the elements considered in the robust design of the drilled shaft in clay. 381
Step 2: Mapping from FORM to MFOSM 382
To establish the mapping function that relates T (computed using MCS or FORM) to 383
M (computed using MFOSM), 18 designs, equally distributed in the design space (DS), are 384
analyzed. The analysis is carried out with respect to both ULS and SLS performance 385
requirements, in terms of 1 and 2, respectively. FORM is adopted herein to evaluate the true 386
reliability indexes of the system performances, as it can be implemented in the Excel 387
spreadsheet environment and yields the results of the similar accuracy as that obtained using 388
MCS. Shown in Figure 7 is an example of FORM analysis of a drilled shaft in clay, with 389
respect to ULS performance, in the spreadsheet environment following the procedures 390
advanced by Low and Tang (2007). Note that for simplicity, the correlation matrix is assumed 391
to retain the original unmodified correlation matrix. Shown in Figure 8 is the reliability 392
analysis of drilled shafts in clay, with respect to both ULS performance and SLS performance, 393
using MFOSM in the spreadsheet environment. As reflected in Figure 9(a) & 9(b), the results 394
are fitted well with a third-order polynomial function, yielding the following mapping 395
functions: 396
M T 3 T 2 T
1 1 1 10.0163( ) 0.1453( ) 0.8862( ) 0.1438 (16) 397
19
M T 3 T 2 T
2 2 2 20.0099( ) 0.1024( ) 0.8164( ) 0.0890 (17) 398
where
and
are the reliability indexes with respect to ULS and SLS, respectively, 399
computed using MFOSM; and, and
are the reliability indexes with respect to ULS and 400
SLS, respectively, computed using FORM. 401
As shown in Figure 9, the clustering is observed at higher reliability index levels (or 402
lower failure probability levels) of drilled shaft performance. Indeed, the reliability analysis of 403
a system with a higher safety level always requires more computational effort and more 404
sophisticated reliability methods; in other words, more computation error might be expected 405
at higher reliability index levels of the system performance due to the less chance of failure 406
(Zhao and Ono 2001). Although higher variability is associated with the reliability analysis 407
results of the drilled shaft at higher reliability index levels, the computed coefficients of 408
determination (R2), with respect to ULS and SLS, are 0.9909 and 0.9839, respectively, 409
indicating the true reliability indexes of the drilled shaft performance can be accurately 410
captured through MFOSM analysis and the corresponding mapping functions. 411
Based upon the obtained mapping functions (see Eqs. 16 & 17), the target reliability 412
indexes (i.e., M
1T and M
2T ) with respect to ULS and SLS, within the context of MFOSM 413
analysis, are set as 2.03 and 1.69, respectively. The resulting target reliability indexes (i.e., 414
M
1T and M
2T ) are readily applicable to the MFOSM-based procedure for evaluating the 415
system performance requirements. Thus, no FORM or MCS is required in the subsequent 416
robust design optimization process, which is a significant feature of the proposed RGD 417
approach that reduces the computational effort. 418
Step 3: Efficient multi-objective optimization 419
20
As formulated previously, the proposed efficient robust geotechnical design (RGD) 420
can be implemented through a series of single-objective optimization problems in which the 421
cost is treated as an additional constraint (see Figure 2b). The single objective optimization-422
based RGD can be efficiently implemented in the spreadsheet environment using Excel’s 423
built-in optimization routine “Solver”, as shown in Figure 10. It is noted that the evaluation of 424
design robustness, in terms of the standard deviation of the performance function with respect 425
to ULS performance, and the evaluation of the system performance requirements, in terms of 426
the reliability indexes with respect to both ULS and SLS, share common computational steps 427
(i.e., they are “coupled” computationally). With the spreadsheet shown in Figure 10, the most 428
robust design within a given cost level (CT) can readily be identified. For example, the most 429
robust designs for three arbitrarily specified cost levels are identified and listed in Table 1. 430
Step 4: Pareto front tradeoff between design robustness and cost efficiency 431
Although the most robust design within a given cost level (CT) can readily be 432
identified using the spreadsheet shown in Figure 10, it should be noted that the optimal cost 433
level is generally unknown in prior, and thus the resulting robust design may be a biased 434
design. To guide the design decision, it is desirable to construct a complete Pareto front that 435
shows the tradeoff between design robustness and cost efficiency. 436
For the given drilled shaft problem, the costs of the least cost and the most robust 437
designs are 2.04 m3 and 2.91 m
3, respectively. The resulting cost interval of [2.04 m
3, 2.91 m
3] 438
is then divided into 10 cost levels, denoted as CT = {2.04 m3, 2.14 m
3, 2.24 m
3, 2.34 m
3, 2.44 439
m3, 2.54 m
3, 2.64 m
3, 2.74 m
3, 2.84 m
3, 2.91 m
3}. For each cost level in CT, the most robust 440
design is identified using the spreadsheet shown in Figure 10. Thereafter, all the resulting 441
optimal designs are plotted as shown in Figure 11. The resulting Pareto front shows a global 442
21
tradeoff between design robustness and cost efficiency in the design space. The Pareto front 443
may be used as a decision aid in selecting the most preferred design at a desired cost level. 444
Also shown in Figure 11 is the Pareto front obtained using the genetic algorithm of 445
NSGA-II (Deb et al. 2002). In NSGA-II, the population size is set at 200 and the generation 446
number is set at 500. At convergence, this multi-objective optimization algorithm (see Figure 447
2a) yields the Pareto front shown in Figure 11. The results show that the proposed RGD 448
approach (see Figure 2b) using Excel can derive practically the same Pareto front as that 449
obtained using more sophisticated multi-objective optimization algorithms such as NSGA-II. 450
As mentioned previously, all the designs on the Pareto front are equally optimal in the 451
sense that no improvement can be achieved in one objective without worsening in the other 452
objective. Nevertheless, if one wishes to select the most preferred design on a Pareto front, the 453
“knee point” concept (Branke et al. 2004; Deb and Gupta 2011) may be used. The knee point 454
on the Pareto front conceptually yields the best compromise among conflicting objectives in 455
the design space. According to Branke et al. (2004), the knee point may be determined using 456
the following marginal utility function: 457
, min , , ( )i j iU' U U i j d d d (18) 458
where U(d, ) is a linear utility function defined as (Branke et al. 2004): 459
, i iU f d d (19) 460
where d are design parameters, in this example: d = (B, D); i is random parameter with a 461
value ranging from 0.0 to 1.0, and i = 1.0; and, fi(d) = objective functions to be minimized, 462
in this case: f1(d) = ([g1] min[g1]) / (max[g1] min[g1]), and f2(d) = (C Cmin) / (Cmax 463
Cmin), where max[g1] and min[g1] represent the maximum value and minimum value of [g1] 464
22
on the Pareto front, respectively; and, Cmax and Cmin represent the maximum value and 465
minimum value of C on the Pareto front, respectively. 466
By means of Monte Carlo simulation (MCS), different random values of 1 and 2 are 467
generated with an assumption of uniform distribution, and the marginal utility function is 468
computed for each of the 10 non-dominated optimal designs on the Pareto front. The design 469
with the maximum expected marginal utility is then taken as the knee point on the Pareto front. 470
Based on the calculated expected marginal utility, Design 2 (see Table 1) is identified as the 471
most preferred solution (or knee point). 472
As can be observed in Figure 11, on the left side of the knee point, a slight reduction 473
in the cost can lead to a large increase in the variation of the ULS performance function 474
(indicating a drastic reduction in the design robustness); on the right side of the knee point, a 475
slight reduction in the variation of the ULS performance function (indicating a slight 476
improvement of design robustness) requires a large increase in cost. Therefore, the knee point 477
on the Pareto front yields the best compromise between design robustness and cost efficiency 478
and can be treated as the most preferred design in the design space 479
480
Further Discussions on the Proposed RGD Approach 481
482
Effectiveness of the variation of the performance function as a robustness measure 483
Figure 12 shows the variation or distribution of the system performance with respect 484
to ULS performance for the three designs listed in Table 1. Through 1,000,000 MCS runs, the 485
distribution of the system performance (defined in Eq. 8) caused by the variation of noise 486
factors can be obtained for each design. The results show that a design with a higher variation 487
of the performance function (i.e., higher robustness) has a wider distribution of system 488
23
performance, which implies a higher degree of uncertainty as to whether the system 489
performance can satisfy the pre-defined performance requirement. 490
Figure 13(a) & 13(b) show the gradient of the system performance to noise factors and 491
the variation of the system performance that arises from noise factors, respectively, for the 492
three designs listed in Table 1. Figure 13(a) depicts that the system performance is most 493
sensitive to the variation of the as-built diameter of drilled shafts (T) and model parameters 494
of QULS (i.e., and Mtc). Figure 13 (b) illustrates that the variation of the system performance is 495
mainly affected by the variation of the normalized undrained shear strength (cn), as-built 496
diameter of drilled shafts (T), and model parameters of QULS (i.e., and Mtc). Compared with 497
the existing robust measures, the presented robustness measure in this paper shows a 498
significant advantage as it is an integration of the gradient (or sensitivity) and the variation of 499
noise factors (see Eq. 2). Thus, the knowledge of how the variation of the system performance 500
is affected by the variation of each noise factor is effectively acknowledged through the 501
formulation of the proposed robustness measure. 502
Figure 14 shows the distribution of the system performance for the three designs listed 503
in Table 1, which is similar to Figure 12, but in Figure 14, the uncertainty in the estimated 504
statistics of noise factors is included. As in most geotechnical problems, it is often challenging 505
to determine the statistics of key geotechnical parameters with certainty; such uncertainty in 506
the estimated statistics can lead to another layer of uncertainty in the system performance. For 507
illustration purposes, the standard deviations of noise factors are treated as independent 508
lognormal variables with a coefficient of variation (COV) of 20%. Through 1,000,000 MCS 509
runs, the distribution of the system performance can be achieved as shown in Figure 14. The 510
results indicate that the design with a lower variation of the performance function (i.e., higher 511
24
robustness) can tolerate a greater uncertainty in the estimated statistics of noise factors, as it 512
yields a shallower distribution of system performance and a lower failure probability, defined 513
as probability of unsatisfactory performance (see Table 2). Therefore, the design with higher 514
design robustness (i.e., a lower variation of the performance function) has an additional 515
benefit of being able to guard against the uncertainty in the estimated statistics of noise factors. 516
Comparison between the RGD and the RBD 517
The Pareto front shown in Figure 11 is employed herein to illustrate the difference 518
between the robust geotechnical design (RGD) and the reliability-based design (RBD). Since 519
all the designs on the Pareto front are feasible designs (i.e., the requirements of system 520
performances are satisfied), the design that yields the least cost (in this case, Design 1) can be 521
readily selected as the final design through the RBD. This least cost design using the RBD 522
approach (Design 1), however, is associated with the highest variation of the system 523
performance (see Figure 12), highest sensitivity of the system performance to the variation of 524
noise factors (see Figure 13), and least tolerance of the system performance to the uncertainty 525
in the estimated statistics of noise factors (see Figure 14). Thus, the least-cost RBD design is 526
not necessarily the most preferred design since a modest variation in the noise factors could 527
lead to a dramatic variation of the system performance. On the other hand, the RGD design 528
approach considers explicitly the safety (performance requirements), cost efficiency, and 529
design robustness, which leads to a Pareto front that offers a tradeoff between design 530
robustness and cost efficiency. Thus, the effect of the uncertain noise factors is realistically 531
managed in the design decision using the proposed RGD approach. 532
Improving design robustness by reducing the variation of noise factors 533
25
So far the discussion focuses on improving the design robustness by adjusting only the 534
design parameters, assuming that the uncertainty in noise factors cannot be reduced or it is not 535
economical to do so. An alternative is to improve the design robustness by reducing the 536
variation of noise factors, referred to as Approach 2 shown in Figure 1. In this section, the 537
construction variation as a noise factor is used as an example to illustrate the robust design 538
using Approach 2. 539
The three designs listed in Table 1 are analyzed for their design robustness, in terms of 540
the variation of the performance function, at three levels of construction variation. Here, the 541
level of construction variation is represented with the standard deviation of as-built geometry 542
(applied to both BT and DT). Three levels of standard deviation, 0.05 m, 0.10 m, and 0.15 m, 543
are adopted for illustration purposes. The results are plotted in Figure 15. As expected, the 544
design robustness can be raised (meaning that the variation of the system performance can be 545
lowered) by reducing the variation of the as-built geometry of drilled shafts that caused by the 546
construction variation. Thus, Approach 2 shown in Figure 1 is demonstrated as an effective 547
alternative for improving the design robustness, although an additional cost that is associated 548
with noise variation reduction may be incurred. On the other hand, cost savings may be 549
achieved from the resulting optimal design at a lower level of construction noise. Although 550
further study to determine the net effect on the overall cost is needed, the above analysis 551
demonstrates that reducing the variation of noise factors might be a cost-effective step in a 552
robust design of a geotechnical system. 553
554
Summary and Conclusions 555
26
This paper presents an efficient robust geotechnical design (RGD) approach, in which 556
the design robustness is represented with the variation of the performance function that is 557
readily available in a reliability analysis. Through the coupling of the evaluation of design 558
robustness with the reliability analysis of system performance, additional savings in the 559
computational effort are achieved. Finally, the multi-objective optimization, typically 560
implemented within the robust design framework, is effectively solved with a series of single 561
objective optimizations using Excel Solver. In short, an efficient and practical robust 562
geotechnical design (RGD) approach is developed. 563
The following conclusions are reached based on the results presented: 564
1) The proposed efficient robust geotechnical design (RGD) approach is shown to 565
be effective and efficient in producing an optimal design that yields the best 566
compromise between design robustness and cost efficiency, while satisfying all 567
design safety and performance requirements. 568
2) The developed robustness measure, in terms of the variation of the performance 569
function, is demonstrated to be effective, intuitive, and fundamentally sound. 570
Higher variation of the performance function signals lower design robustness, 571
which implies a higher degree of uncertainty as to whether the system can satisfy 572
the pre-defined performance requirement. 573
3) Within the context of the proposed RGD approach, the evaluation of design 574
robustness and the evaluation of system performance, two essential procedures 575
in the robust design, share computational steps, as both can be analyzed using 576
MFOSM with a common formulation. Thus, the computational efficiency is 577
greatly improved over other existing reliability-based robust design approaches. 578
27
4) Although the RGD is presented as a multi-objective optimization problem, it can 579
be efficiently solved using a series of single-objective optimizations that does 580
not require the more sophisticated genetic algorithms and programming skills. 581
The Pareto front obtained from a series of single-objective optimizations 582
implemented with Excel Solver is shown to be identical to the one obtained with 583
the more sophisticated multi-objective optimization algorithm. 584
5) While the robust design using Approach 1 as shown in Figure 1 has been shown 585
effective, it is possible that Approach 2, which allows for consideration of some 586
reduction in the variation of noise factors within the framework of the robust 587
design, can yield more cost-efficient designs while improving the design 588
robustness. 589
6) As shown in Figure 6, the proposed RGD approach considers the noises from 590
construction variation, loading conditions, model errors, and geotechnical 591
parameters in the design of drilled shafts in clay. The proposed approach is 592
flexible in allowing use of either Approach 1 or Approach 2 or both for the more 593
preferred design with respect to cost efficiency, design robustness, and 594
performance requirements. In light of the recognition that the capacity and 595
performance of a drilled shaft are strongly affected by the construction quality, 596
the proposed RGD approach offers a practical and effective tool to examine 597
possible alternatives. 598
7) While the proposed efficient RGD approach is demonstrated only through an 599
example of a drilled shaft design, the formulation and solution algorithm are set 600
up in a way that is suitable for general applications to most engineering problems. 601
28
Thus, the proposed RGD approach can be easily adapted for design of other 602
geotechnical systems. 603
604
Acknowledgments 605
The study on which this paper is based was supported in part by National Science 606
Foundation through Grant CMMI-1200117 (“Transforming Robust Design Concept into a 607
Novel Geotechnical Design Tool”) and the Glenn Department of Civil Engineering, Clemson 608
University. The results and opinions expressed in this paper do not necessarily reflect the 609
views and policies of the National Science Foundation. 610
29
References 611
Ang, A. H., and Tang, W. H. (2004). Probability Concepts in Engineering (2nd edition). John 612
Wiley and Sons, New York. 613
Baecher, G. B., and Christian, J. T. (2003). Reliability and Statistics in Geotechnical 614
Engineering. John Wiley and Sons, New York. 615
Beyer, H. G., and Sendhoff, B. (2007). “Robust optimization – A comprehensive survey.” 616
Computer Methods in Applied Mechanics and Engineering, 196(33), 3190-3218. 617
Branke, J., Deb, K., Dierolf, H., and Osswald, M. (2004). “Finding knees in multi-objective 618
optimization.” Parallel Problem Solving from Nature-PPSN VIII, 722-731. 619
Chen, W., Allen, J. K., Tsui, K. L., and Mistree, F. (1996). “Procedure for robust design: 620
minimizing variations caused by noise factors and control factors.” Journal of 621
Mechanical design, Transactions of the ASME, 118(4), 478-485. 622
Christian, J. T., Ladd, C. C., and Baecher, G. B. (1994). “Reliability applied to slope stability 623
analysis.” Journal of Geotechnical Engineering, 120(12), 2180-2207. 624
Deb, K., Pratap, A., Agarwal, S., and Meyarivan, T. (2002). “A fast and elitist multi-objective 625
genetic algorithm: NSGA-II.” IEEE Transactions on Evolutionary Computation, 6(2), 626
182-197. 627
Deb, K., and Gupta, S. (2011). “Understanding knee points in bicriteria problems and their 628
implications as preferred solution principles.” Engineering Optimization, 43(11), 629
1175-1204. 630
Dithinde, M., Phoon, K. K., De Wet, M., and Retief, J. V. (2010). “Characterization of model 631
uncertainty in the static pile design formula.” Journal of Geotechnical and 632
Geoenvironmental Engineering, 137(1), 70-85. 633
30
Doltsinis, I., Kang, Z., and Cheng, G. (2005). “Robust design of non-linear structures using 634
optimization methods.” Computer Methods in Applied Mechanics and Engineering, 635
194(12), 1779-1795. 636
Du, X., and Chen, W. (2000). “Towards a better understanding of modeling feasibility 637
robustness in engineering design.” Journal of Mechanical Design, 122, 385-394. 638
ETC10 (2009). Evaluation of Eurocode 7: Design Examples 2, 639
http://www.eurocode7.com/etc10/. 640
Faravelli, L. (1989). “Response-surface approach for reliability analysis.” Journal of 641
Engineering Mechanics, 115(12), 2763-2781. 642
Fenton, G. A., and Griffiths, D. V. (2008). Risk Assessment in Geotechnical Engineering. 643
John Wiley & Sons, New York 644
Gong, W., Wang, L., Juang, C. H., Zhang, J., and Huang, H. (2014). “Robust geotechnical 645
design of shield-driven tunnels.” Computer and Geotechnics, 646
http://dx.doi.org/10.1016/j.compgeo.2013.12.006. 647
Haimes, Y. Y., Lasdon, L. S., and Wismer, D. A. (1971). “On a bicriterion formulation of the 648
problems of integrated system identification and system optimization.” IEEE 649
Transactions on Systems, Man, and Cybernetics, 1(3), 296-297. 650
Iskander, M., Roy, D., Kelley, S., and Ealy, C. (2003). “Drilled shaft defects: Detection, and 651
effects on capacity in varved clay.” Journal of Geotechnical and Geoenvironmental 652
Engineering, 129(12), 1128-1137. 653
Juang, C. H., and Wang, L. (2013) “Reliability-based robust geotechnical design of spread 654
foundations using multi-objective genetic algorithm” Computers and Geotechnics, 48, 655
96-106. 656
31
Juang, C. H., Wang, L., Liu, Z., Ravichandran, N., Huang, H., and Zhang, J. (2013a). “Robust 657
geotechnical design of drilled shafts in sand: New design perspective.” Journal of 658
Geotechnical and Geoenvironmental Engineering, 139(12), 2007-2019. 659
Juang, C. H., Wang, L., Khoshnevisan, S., and Atamturktur, S. (2013b). “Robust geotechnical 660
design – methodology and applications.” Journal of GeoEngineering, Taiwan 661
Geotechnical Society, 8(3), 61-70. 662
Juang, C. H., Wang, L., Hsieh, H. S., and Atamturktur, S. (2013c), “Robust geotechnical 663
design of braced excavations in clays,” Structural Safety (in press). 664
Kim, I. Y., and De Weck, O. L. (2005). “Adaptive weighted-sum method for bi-objective 665
optimization: Pareto front generation.” Structural and Multidisciplinary Optimization, 666
29(2), 149-158. 667
Lee, J. H., Kim, T., and Lee, H. (2010). “A study on robust indentation techniques to evaluate 668
elastic–plastic properties of metals.” International Journal of Solids and Structures, 669
47(5), 647-664. 670
Li, D., Chen, Y., Lu, W., and Zhou, C. (2011). “Stochastic response surface method for 671
reliability analysis of rock slopes involving correlated non-normal variables.” 672
Computers and Geotechnics, 38(1), 58-68. 673
Liu Z., Atamturktur S., and Juang H., (2013). “Performance Based Robust Design 674
Optimization of Steel Moment Resisting Frames,” Journal of Constructional Steel 675
Research, 89, 165-174. 676
Low, B. K., and Tang, W. H. (2007). “Efficient spreadsheet algorithm for first-order 677
reliability method.” Journal of Engineering Mechanics, 133(12), 1378-1387. 678
32
Marano, G. C., Sgobba, S., Greco, R., and Mezzina, M. (2008). “Robust optimum design of 679
tuned mass dampers devices in random vibrations mitigation.” Journal of Sound and 680
Vibration, 313(3), 472-492. 681
Marler, R. T., and Arora, J. S. (2010). “The weighted sum method for multi-objective 682
optimization: new insights.” Structural and Multidisciplinary Optimization, 41(6), 683
853-862. 684
Nagy, Z. K., and Braatz, R. D. (2004). “Open-loop and closed-loop robust optimal control of 685
batch processes using distributional and worst-case analysis.” Journal of Process 686
Control, 14(4), 411-422. 687
O’Neill, M. W. (2001). “Side resistance in piles and drilled shafts.” 34th
Terzaghi Lecture, 688
Journal of Geotechnical and Geoenvironmental Engineering, 127(1), 1-16. 689
Orr, T. L. L., Bond, A. J. and Scarpelli, G. (2011). “Findings from the 2nd Set of Eurocode 7 690
Design Examples.” Materiały konferencji ISGSR (Red.: Vogt, Schuppener. Straub i 691
Bräu), Bundesanstalt für Wasserbau, Munich, Germany. 692
Park, G. J., Lee, T. H., Lee, K. H., and Hwang, K. H. (2006). “Robust design: an overview.” 693
AIAA Journal, 44(1), 181-191. 694
Parkinson, J. S. (1993). “Signal transduction schemes of bacteria.” Cell, 73(5), 857-871. 695
Phadke, M. S. (1989) Quality Engineering Using Robust Design. Prentice Hall. Englewood 696
Cliffs, New Jersey. 697
Phoon, K. K., and Kulhawy, F. H. (2005). “Characterization of model uncertainties for drilled 698
shafts under undrained axial loading.” Contemporary Issues in Foundation 699
Engineering, Geotechnical Special Publication (GSP) No. 131, ASCE, Reston, VA. 700
33
Phoon, K. K., Kulhawy, F. H., and Grigoriu, M. D. (2003). “Development of a reliability-701
based design framework for transmission line structure foundations.” Journal of 702
Geotechnical and Geoenvironmental Engineering, 129(9), 798-806. 703
Poulos, H. G. (2005). “Pile behavior–Consequences of geological and construction 704
imperfections.” Journal of Geotechnical and Geoenvironmental Engineering, 131(5), 705
538-563. 706
Hämäläinen, R. P., and Mäntysaari, J. (2002). “Dynamic multi-objective heating optimization.” 707
European Journal of Operational Research, 142(1), 1-15. 708
Taguchi, G. (1986). Introduction to quality engineering: designing quality into products and 709
processes. Quality Resources, White Plains, New York. 710
Tsui, K. L. (1992). “An overview of Taguchi method and newly developed statistical methods 711
for robust design.” IIE Transactions, 24(5), 44-57. 712
Wang, J., Wang, Y., and Cao, Z. (2011) “A Comparative Study of Pile Design Using 713
Eurocode 7 and RBDE.” Proceedings, 3rd International Symposium on Geotechnical 714
Safety and Risk (ISGSR2011), 481-487. 715
Wang, L., Hwang, J. H., and Juang, C. H., and Sez Atamturktur (2013), “Reliability-based 716
design of rock slopes – A new perspective on design robustness,” Engineering 717
Geology, 154, 56-63. 718
Wu, T. H., Tang W. H., Sangrey, D. A., and Baecher, G. B. (1989). “Reliability of offshore 719
foundations – State-of-the-art.” Journal of Geotechnical Engineering, 115(2), 157-178. 720
Zhang, J., Zhang, L. M., and Tang, W. H. (2011). “Reliability-based optimization of 721
geotechnical systems.” Journal of Geotechnical and Geoenvironmental Engineering, 722
137(12), 1211-1221. 723
34
Zhang, W., Chen, B. S., and Tseng, C. S. (2005). “Robust H filtering for nonlinear stochastic 724
systems.” Signal Processing, IEEE Transactions on, 53(2), 589-598. 725
Zhao, Y. G., and Ono, T. (2001). “Moment methods for structural reliability.” Structural 726
safety, 23(1), 47-75. 727
Zhu, S., and Fukushima, M. (2009). “Worst-case conditional value-at-risk with application to 728
robust portfolio management.” Operations Research, 57(5), 1155-1168. 729
730
35
Lists of Tables
Table 1. Examples of most robust design at various cost levels
Table 2. Probability of unsatisfactory performance of each of the three designs at each of the
three levels of variation of noise factors
36
Table 1. Examples of most robust design at various cost levels
Specified
cost level,
CT (m3)
Most robust design
within the cost level
Design parameters Actual
cost,
C (m3)
Resulting
robustness,
[g1(d, θ)] (kN)
Diameter,
B (m)
Depth,
D (m)
2.04 Design 1 0.32 25.00 2.04 800.46
2.24 Design 2 0.42 16.19 2.24 476.26
2.91 Design 3 0.60 10.29 2.91 347.27
37
Table 2. Probability of unsatisfactory performance of each of the three designs at each of the
three levels of variation of noise factors
Design
Level of variation of noise factors (in terms of COV of
the standard deviation of each of noise factors)
COV = 10% COV = 20% COV = 30%
Design 1 1.169 10-3
2.135 10-3
3.937 10-3
Design 2 0.949 10-3
1.667 10-3
3.044 10-3
Design 3 0.530 10-3
0.917 10-3
1.843 10-3
38
Lists of Figures
Figure 1. Illustration of possible approaches to improve design robustness
Figure 2. Optimization setting of robust design
(a) Robust design through multi-objective optimization
(b) Robust design through proposed optimization procedure
Figure 3. Illustration of non-dominated solution and Pareto in a robust design optimization
(note: lower [g] signals higher design robustness)
Figure 4. Schematic diagram of a drilled shaft in clay (after ETC 10)
Figure 5. Characterization of undrained shear strength from test data
Figure 6. Illustration of robust design of a drilled shaft in clay
Figure 7. Layout of a spreadsheet for FORM analysis of a drilled shaft based on ULS
Figure 8. Layout of a spreadsheet for MFOSM analysis of a drilled shaft based on ULS and
SLS
Figure 9. Mapping from FORM to MFOSM
(a) Mapping function of reliability index based on ULS
(b) Mapping function of reliability index based on SLS
Figure 10. Layout of a spreadsheet for robust design of a drilled shaft
Figure 11. Pareto front showing tradeoff between design robustness and cost efficiency (note:
lower [g1(d, )] signals higher design robustness)
Figure 12. Distribution of ULS performance of three drilled shaft designs (standard deviation
of each and every noise factor is a constant)
Figure 13. Sensitivity of performance function with respect to noise factors
(a) Gradient of performance function to each noise factor
(b) Effect of each noise factor on the variation of performance function
Figure 14. Distribution of ULS performance of three drilled shaft designs (standard deviation
of each and every noise factor is a random variable with 20% COV)
Figure 15. Robustness of three designs at various levels of construction variation (in terms of
standard deviation of as-built geometry)
39
Table 1. Examples of most robust design at various cost levels
Specified
cost level,
CT (m3)
Most robust design
within the cost level
Design parameters Actual
cost,
C (m3)
Resulting
robustness,
[g1(d, θ)] (kN)
Diameter,
B (m)
Depth,
D (m)
2.04 Design 1 0.32 25.00 2.04 800.46
2.24 Design 2 0.42 16.19 2.24 476.26
2.91 Design 3 0.60 10.29 2.91 347.27
40
Table 2. Probability of unsatisfactory performance of each of the three designs at each of the
three levels of variation of noise factors
Design
Level of variation of noise factors (in terms of COV of
the standard deviation of each of noise factors)
COV = 10% COV = 20% COV = 30%
Design 1 1.169 10-3
2.135 10-3
3.937 10-3
Design 2 0.949 10-3
1.667 10-3
3.044 10-3
Design 3 0.530 10-3
0.917 10-3
1.843 10-3
41
Figure 1. Illustration of possible approaches to improve design robustness
Noise factors, 1
Sy
stem
res
po
nse
, g
(d1,
1)
Robust design Noise factors, 1
Sy
stem
res
po
nse
, g
(d2,
1)
Noise factors, 2
Sy
stem
res
po
nse
, g
(d1,
2)
Approach 2
Approach 1
42
(a) Robust design through multi-objective optimization
(b) Robust design through proposed optimization procedure
Figure 2. Optimization setting of robust design
Find: d (Design parameters)
Subjected to: d DS (Design space)
M
i M
iT (Requirements of system performances)
Objectives: min [g] (Maximizing the design robustness)
min C (Minimizing the cost)
Find: d (Design parameters)
Subjected to: d DS (Design space)
M
i M
iT (Requirements of system performances)
C < CTj (Cost level is dealt as an additional constraint)
Objectives: min [g] (Maximizing the design robustness)
43
Figure 3. Illustration of non-dominated solution and Pareto front in a robust design
optimization (note: lower [g] signals higher design robustness)
(CTj,
j[g])
Infeasible domain
Pareto front
Ro
bu
stn
ess
mea
sure
,
[g]
Cost, C
Feasible domain
Non-dominated solution
44
Figure 4. Schematic diagram of a drilled shaft in clay (after ETC 10)
Fd + Fl
D = ?
B = ?
Stiff clay
= 20 kN/m3
cu is obtained from
undrained triaxial test
sa = 20 mm
45
Figure 5. Characterization of undrained shear strength from test data
30
25
20
15
10
5
0 0 100 200 300 400 500 600
E[cu] [cu]
E[cu] + 2[cu]
Data from test
Undrained shear strength, cu (kPa)
Dep
th,
z (m
)
E[cu]
46
Figure 6. Illustration of robust design of a drilled shaft in clay
Robust design
of drilled shafts in clay
Design objectives
Maximize design robustness
Minimize cost
Satisfy design constraints
Noise on
geotechnical parameters
(cn)
Noise on
model errors
(, Mtc, a, and b)
Noise on
loading conditions
(Fd, Fl)
Noise on
as-built geometry
(BT and DT)
47
Figure 7. Layout of a spreadsheet for FORM analysis of a drilled shaft based on ULS
Distribution n n n *
cn (kPa/m) Lognormal 13.04 4.15 2.520 0.311 -1.008 9.09 B (m) 0.45
B T (m) Lognormal 0.45 0.05 -0.805 0.111 -0.893 0.41 D (m) 10
D T (m) Lognormal 10.00 0.05 2.303 0.005 -0.478 9.98
Normal 0.00 0.12 0.000 0.120 -1.141 -0.14
Mtc Lognormal 0.45 0.13 -0.839 0.283 -0.439 0.38 g 1(d , * 0.00
a (mm) Lognormal 2.79 2.04 0.812 0.654 0.000 2.25
b Lognormal 0.82 0.09 -0.204 0.109 0.000 0.82
F d (kN) Lognormal 258.00 25.80 5.548 0.100 0.442 268.29
F l (kN) Lognormal 116.00 20.88 4.738 0.179 0.359 121.72
1 (ULS) 1.906
PT
f 1 (ULS) 2.83E-02
1 0 0 0 0 0 0 0 0
0 1 0.5 0 0 0 0 0 0
0 0.5 1 0 0 0 0 0 0
0 0 0 1 0 0 0 0 0
0 0 0 0 1 0 0 0 0
0 0 0 0 0 1 -0.8 0 0
0 0 0 0 0 -0.8 1 0 0
0 0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 0 1
Notes:
For this case, in solver's option, use automatic scaling
Quadratic for Estimates, Central for Derivatives and Newton for Search
Others as default options.
Initially, enter the mean , standard deviation , and correlation matrix [R ] of noise factors and the geometry of drilled shaft (in terms of B
and D ), followed by invoking Excel Solver to automatically minimize reliability index by changing column n , subjected to g 1(d , ) = 0.
Correlation matrix [R ]
Results
n i = 1
[F ( i )]Statistical characterization of noise
factors
Equivalent normal
parameters
Performance function
Design parameters (d )
1
T 1
1( , *) 0
T T
1 1
min [ ] [ ] [ ]
( )
T
g
f
R
P
dn n
See Eq. (8)
48
Figure 8. Layout of a spreadsheet for MFOSM analysis of a drilled shaft based on ULS and SLS
g d , i i g d , i i
cn (kPa/m) 13.04 4.15 37.94 39.54 E [g 1 (d , )] B (m) 0.45
T (m) 0.45 0.05 2033.75 2119.59 329.66 D (m) 10.00
D T (m) 10.00 0.05 102.94 107.28 [g 1 (d , )]
0.00 0.12 921.72 960.62 228.62
Mtc 0.45 0.13 438.53 457.04
1 (ULS) 1.442
a (mm) 2.79 2.04 0.00 -37.81 E [g 2 (d , )]
2 (SLS) 1.481
b 0.82 0.09 0.00 -757.21 359.36 PM
f 1 (ULS) 7.47E-02
F d (kN) 258.00 25.80 -1.00 -1.00 [g 2 (d , )] PM
f 2 (SLS) 6.93E-02
F l (kN) 116.00 20.88 -1.00 -1.00 242.6106849
Covariance Matrix [C ]
17.223 0 0 0 0 0 0 0 0
0 0.003 0.001 0 0 0 0 0 0
0 0.001 0.003 0 0 0 0 0 0
0 0 0 0.014 0 0 0 0 0
0 0 0 0 0.017 0 0 0 0
0 0 0 0 0 4.162 -0.147 0 0
0 0 0 0 0 -0.147 0.008 0 0
0 0 0 0 0 0 0 665.640 0
0 0 0 0 0 0 0 0 435.974
Notes:
The partial derivative of performance function is approximated as g / i =Dg /D i , where D i = 0.1 i
Initially, enter the mean , standard deviation , and covariance matrix [C ] of noise factors and the geometry of drilled shaft (in terms of B and D ),
followed by automatocal computation of the reliability indexes (in terms of
1 and M
2 ).
Partial derivatives of performance
fucntions G
Statistics characterization
of noise factors
Statistics of performance
functions g 1 and g 2
Design parameters (d )
Results
ULS
behavior
SLS
behavior
See Eq. (5)
See Eqs. (2) & (4)
49
(a) Mapping function of reliability index based on ULS
(b) Mapping function of reliability index based on SLS
Figure 9. Mapping from FORM to MFOSM
0 1 2 3 4 50.0
0.5
1.0
1.5
2.0
2.5
3.0
M 1
fro
m M
FO
SM
T
1 from FORM
Eq. (16), R2 = 0.9909
0 1 2 3 4 50.0
0.5
1.0
1.5
2.0
2.5
3.0
M 2
fro
m M
FO
SM
T
2 from FORM
Eq. (17), R2 = 0.9839
50
Figure 10. Layout of a spreadsheet for robust design of a drilled shaft
g d , i i g d , i i
cn (kPa/m) 13.04 4.15 78.34 81.65 E [g 1 (d , )] Cost level C T (m
3) 2.24
T (m) 0.42 0.05 3932.32 4098.30 966.81
T 1 (ULS) 2.03
D T (m) 16.19 0.05 130.78 136.30 [g 1 (d , )]
T 2 (SLS) 1.69
0.00 0.12 2253.12 2348.22 476.26
Mtc 0.45 0.13 640.33 667.36
a (mm) 2.79 2.04 0.00 -72.05 E [g 2 (d , )]
b 0.82 0.09 0.00 -1442.85 1023.40
F d (kN) 258.00 25.80 -1.00 -1.00 [g 2 (d , )] B (m) 0.42
F l (kN) 116.00 20.88 -1.00 -1.00 504.19184 D (m) 16.19
1 (ULS) 2.03
Covariance Matrix [C ]
2 (SLS) 2.03
17.223 0 0 0 0 0 0 0 0
0 0.003 0.001 0 0 0 0 0 0 PM
f 2 (SLS) 2.12E-02
0 0.001 0.003 0 0 0 0 0 0 Cost C (m3) 2.24
0 0 0 0.014 0 0 0 0 0 [g 1 (d , )] 476.26
0 0 0 0 0.017 0 0 0 0
0 0 0 0 0 4.162 -0.147 0 0
0 0 0 0 0 -0.147 0.008 0 0
0 0 0 0 0 0 0 665.640 0
0 0 0 0 0 0 0 0 435.974
Notes:
The partial derivative of performance function is approximated as g / i =Dg /D i , where D i = 0.1 i
For this case, in solver's option, use automatic scaling
Quadratic for Estimates, Central for Derivatives and Newton for Search
Others as default options.
Initially, enter the mean , standard deviation , and covariance matrix [C ] of noise factors , the geometry of drilled shaft (in terms of B and D ), and the cost level C T ,
followed by invoking Excel Solver to automatically minimize the standard deviation of performance fucntion [g 1(d , )] by changing design parameters B and D ,
subjected to Bm, 0.6m], D[10.0m, 25.0m], M
1 > M
T 1, M
2 > M
T 2, and C < C T
Partial derivatives of performance
fucntions G
Statistics characterization
of noise factors
Statistics of performance
functions g 1 and g 2
Robust optimizatin constraint
ULS
behavior
SLS
behavior
Performance
requirements
Most robust design within given cost level
Design
parameters
Objectives
2.12E-02
Product
performance PM
f 1 (ULS)
See Eqs. (2) & (4) M
T Tf
51
Figure 11. Pareto front showing tradeoff between design robustness and cost efficiency
(note: lower [g1(d, )] signals higher design robustness)
2.0 2.2 2.4 2.6 2.8 3.00
200
400
600
800
1000
NSGA-II solution
Robust
nes
s m
easu
re,
[g1(d
,)]
(kN
)
Cost, C (m3)
Design 1
Design 2
(knee point) Design 3
Spreadsheet solution
52
Figure 12. Distribution of ULS performance of three drilled shaft designs
(standard deviation of each and every noise factor is a constant)
-1000 0 1000 2000 3000 4000 50000.0000
0.0003
0.0006
0.0009
0.0012
0.0015
Design 3
Design 2
(knee point)
Pro
bab
ilit
y d
ensi
ty f
un
ctio
n,
PD
F
Performance function of ULS, g1 (kN)
Design 1
53
(a) Gradient of performance function to each noise factor
(b) Effect of each noise factor on the variation of performance function
Figure 13. Sensitivity of performance function with respect to noise factors
-2000
0
2000
4000
6000
8000
baMTc
Fl
DT
BT
Fd
Design 1
Design 2 (knee point)
Design 3
Gra
die
nt
to n
ois
e fa
cto
rs
cn
-200
0
200
400
600
800
baMTc
Fl
DT
BT
Fd
Design 1
Design 2 (knee point)
Design 3
Var
iati
on
fro
m n
ois
e fa
cto
rs
cn
54
Figure 14. Distribution of ULS performance of three drilled shaft designs
(standard deviation of each and every noise factor is a random variable with 20% COV)
-1000 0 1000 2000 3000 4000 50000.0000
0.0003
0.0006
0.0009
0.0012
0.0015
Design 3
Design 2
(knee point)
Pro
bab
ilit
y d
ensi
ty f
un
ctio
n,
PD
F
Performance fucntion of ULS, g1 (kN)
Design 1
55
Figure 15. Robustness of three designs at various levels of construction variation (in terms of
standard deviation of as-built geometry)
0.00 0.05 0.10 0.15 0.200
300
600
900
1200
1500
Robust
nes
s m
easu
re,
[g1(d
,)]
(kN
)
Standard deviation of as-built geometry (m)
Design 1
Design 2 (knee point)
Design 3