Sie sind auf Seite 1von 23

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

Towards reliability-based design for geotechnical engineering


K. K. Phoon National University of Singapore, Singapore

ABSTRACT: This paper presents an overview of the evolution in structural and geotechnical design practice over the past half a decade or so in relation to how uncertainties are dealt with. For the general reader who is encountering reliability-based design (RBD) for the first time, this would provide a valuable historical perspective of our present status and important outstanding issues that remain to be resolved. The key elements of RBD are briefly discussed and the availability of statistics to provide empirical support for the development of simplified RBD equations is highlighted. Several important implementation issues are presented with reference to an EPRI study for reliability-based design of transmission line structure foundations. Reliability-based design, simplified or otherwise, provides a more consistent means of managing uncertainties, but it is by no means a perfect solution. Engineering judgment still is indispensable in many aspects of geotechnical engineering reliability analysis merely removes the need for guesswork on how uncertainties affect performance and is comparable to the use of elasto-plastic theory to remove the guesswork on how loads induce stresses and deformations. INTRODUCTION The basis for making a geotechnical design decision is not as well studied nor subjected to the same degree of formal scrutiny as structural design. Goble (1999) noted in an NCHRP Synthesis Report on Geotechnical Related Development and Implementation of Load and Resistance Factor Design (LRFD) Methods that the "education of geotechnical engineers strongly emphasizes the evaluation of soil and rock properties" and "the design process does not receive the emphasis that it does in structural engineering education". In fact, the first time the topic "Codes and Standards" was selected for formal discussion in a major ISSMGE conference was in 1989 (Ovesen 1989). Examination of current practice shows that procedures for selecting nominal soil strengths are not well-defined or followed uniformly. Some engineers use the mean value, while others use the most conservative of the measured strengths (Whitman, 1984). Different calculation methods are preferred in different localities or even by different engineers in the same locality (Goble 1999). The manner in which the factor of safety is incorporated in the design equation also is highly varied (Kulhawy 1984, 1996). Green & Becker (2001) made similar observations in a National Report on Limit State Design in Geotechnical Engineering in Canada. Golder (1966) noted quite aptly in a discussion of the second Terzaghi Lecture by Arthur Casagrande that: "We do not know how we make a decision." Ironically, this view is rarely acknowledged publicly in the geotechnical engineering profession. The predominant view ranges from "If it is not broke, why fix it" (Green & Becker 2001) to a general feeling that conventional practice is perfectly adequate to do optimal design (Committee on Reliability Methods 1995, Kulhawy 1996). Although existing practice has undoubtedly served the profession well for many years, this paper argues that genuine improvements are possible if our practice were to be complemented by reliability-based design (RBD) methodologies. No one is advocating total abandonment of existing practice for something entirely new. In fact, the reverse is probably closer to the truth - many aspects of current design practice would still appear in new RBD codes, albeit in a modified form (Kulhawy & Phoon 1996). It is also important to highlight that geotechnical design would be subjected to increasing codification as a result of code harmonization across material types and national boundaries. It is also 1

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

clear that regulatory pressure eventually would bring geotechnical design within an umbrella framework predominantly established by structural engineers. In the United States, this process is already well underway for highway bridge design. In reference to AASHTO LRFD Specifications (AASHTO 2002), Withiam (2003) noted that the leverage to drive implementation was the intention to sunset (i.e. no longer publish) the long-standing AASHTO Standard Specifications that have provided national requirements for highway bridge superstructure and substructure design since the 1930s. The difficulty of maintaining status quo was highlighted about 10 years ago in the National Research Council Report on Reliability Methods for Risk Mitigation in Geotechnical Engineering (Committee on Reliability Methods 1995). A fundamental change in mindset, similar to what has taken place in the structural community since the 1970s, is needed for the profession to take the next step. It is accurate to say that this change has not taken place in the geotechnical community in North America (Goble 1999, DiMaggio et al. 1999, Green & Becker 2001), although significant initiatives have been launched by major agencies in recent years such as OHBDC3 (Ministry of Transportation Ontario 1992), CAN/CSA-S472-92 (CSA 1992a), API RP 2A-LRFD (API 1993), EPRI (Phoon et al. 1995), NBCC (National Research Council of Canada 1995), AASHTO LRFD Bridge Code (AASHTO 2002), and CHBDC (CSA 2000). Currently, the geotechnical community is mainly preoccupied with the transition from working or allowable stress design (WSD/ASD) to Load and Resistance Factor Design (LRFD). The term "LRFD" is used in a loose way to encompass methods that require all limit states to be checked using a specific multiple-factor format involving load and resistance factors. This term is used most widely in the United States and is equivalent to "Limit State Design (LSD)" in Canada. Both LRFD and LSD are philosophically akin to the partial factors approach commonly used in Europe, although a different multiple-factor format involving factored soil parameters is used. The emphasis in LRFD or its equivalent in Canada and Europe is primarily on the re-distribution of the original global factor safety in WSD into separate load and resistance factors (or soil parameter partial factors). The absence of strong analytical calibration and verification in Eurocode 7 (CEN/TC250 1994) and OHBDC3 (Ministry of Transportation Ontario 1992) is noted by DiMaggio et al. (1999) in an FHWA Report on "Geotechnical Engineering Practices in Canada and Europe". Paikowsky & Stenersen (2000) also noted a similar lack of data supporting current AASHTO LRFD specifications. There are strong practical reasons to consider geotechnical LRFD as a simplified reliability-based design procedure, rather than an exercise in rearranging the original global factor of safety. This calls for a willingness to accept the fundamental philosophy that: (a) absolute reliability is an unattainable goal in the presence of uncertainty and (b) probability theory can provide a formal framework for developing design criteria that would ensure that the probability of "failure" (used herein to refer to exceeding of any prescribed limit state) is acceptably small. In other words, geotechnical LRFD should be derived as the logical end-product of a philosophical shift in mindset to probabilistic design in the first instance and a simplification of rigorous reliability-based design into a familiar look and feel design format in the second. The need to draw a clear distinction between accepting reliability analysis as a necessary theoretical basis for geotechnical design and downstream calibration of simplified multiple-factor design formats, with emphasis on the former, was emphasized by Phoon et al. (2003a). The former provides a consistent method for propagation of uncertainties and a unifying framework for risk assessment across disciplines (structural and geotechnical design) and national boundaries. Other competing frameworks have been suggested (e.g., -method by Simpson et al. 1981, worst attainable value method by Bolton 1989, Taylor series method by Duncan 2000) but none has the theoretical breadth and power to handle complex real-world problems that may require nonlinear 3-D finite element or other numerical approaches for solution. In the development of Eurocode 7, much attention has been focused on the geotechnical aspects of code harmonisation (Frank 2002, Ovesen 2002, Orr 2002). This clearly takes precedence over safety aspects, but the time is perhaps ripe to decide if a theoretical platform is necessary to rationalise risk assessment. If the platform is not reliability analysis, then what alternative is available? The need to derive simplified reliability-based design (RBD) equations perhaps is of practical importance to maintain continuity with past practice, but it is not necessary and increasingly fraught with 2

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

difficulties when sufficiently complex problems are posed. The limitations faced by simplified RBD have no bearing on the generality of reliability theory. This is analogous to arguing that limitations in closedform elastic solutions are related to elasto-plastic theory. The use of finite element software on relatively inexpensive and powerful PCs (with gigahertz processors, gigabyte of memory, and hundreds of gigabytes - verging on terabyte - of disk) permit real world problems to be simulated on an unprecedented realistic setting almost routinely. Phoon et al. (2003a) presented some examples to clarify that limitations of the implementation (say LRFD) do not carry over to the underlying reliability framework. Attention should be focused on the more basic issue pertaining to the relevance of reliability theory in geotechnical design. This paper presents an overview of the evolution in structural and geotechnical design practice over the past half a decade or so in relation to how uncertainties are dealt with. For the general reader who is encountering reliability-based design for the first time, this would provide a valuable historical perspective of our present status and important outstanding issues that remain to be resolved. The key elements of RBD are briefly discussed and the availability of statistics to provide empirical support for the development of simplified RBD equations is highlighted. Several important implementation issues are presented with reference to an EPRI study for reliability-based design of transmission line structure foundations (Phoon et al. 1995). Numerous examples of RBD calibration are given elsewhere (Phoon & Kulhawy 2002a, b; Phoon et al. 2003c; Phoon & Kulhawy 2004) HISTORICAL OVERVIEW Structural LRFD The classical structural reliability theory became widely known through a few influential publications such as Freudenthal (1947) and Pugsley (1955). The fundamental philosophy is that absolute reliability is an unattainable goal in the presence of uncertainty. Probability theory can provide a formal framework for developing design criteria that would ensure that the probability of "failure" (used herein to refer to exceeding of any prescribed limit state) is acceptably small. While the philosophy is elegant, the theory is mathematically intractable and numerically cumbersome. Cornell (1969) probably was the first to introduce the concept of a reliability index for simplified probabilistic design. Only second-moment information (mean and covariance) on uncertain parameters was needed and the computation was made simple by adopting the Gaussian model for random variables. However, the idea still was rather radical and could have been ignored if not for Lind (1971), who demonstrated that Cornell's reliability index could be used to derive load and resistance factors formally. The ability to repackage probabilistic design into a simplified multiple-factor design format with the same look and feel as existing design formats, while retaining theoretical rigor, is an important development from a practical point of view. To the authors' knowledge, LRFD was first implemented for steel building structures by Ravindra & Galambos (1978) using the theoretical basis established by Cornell (1969) and Lind (1971). In the meantime, serious theoretical difficulties were encountered with Cornell's index, with the most severe being the problem of invariance. Cornell's index was found to vary when certain simple limit states were reformulated in a mechanically equivalent way. Although second-moment reliability-based structural design was becoming widely accepted in the early seventies, the goal of developing simplified design criteria firmly founded on a rigorous reliability basis remained elusive. This unsatisfactory condition was resolved eventually by Hasofer & Lind (1974), when they proved mathematically that the nearest distance of the limit state function from the origin of a standard Gaussian space is an invariant measure of reliability. This major theoretical breakthrough enforces invariance while retaining the practical second-moment simplification of Cornell's index. The last piece of significant addition to the theoretical repertoire for solving time-invariant reliability problems was the algorithm of Rackwitz & Fiessler (1978), which provided a practical and computationally efficient recipe for computing this reliability index with no restriction on the number of random variables. The reliability method proposed by Hasofer & Lind (1974) and its subsequent generalizations to handle non-Gaussian and correlated 3

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

random variables commonly is called the First-Order Reliability Method (FORM). Ellingwood et al. (1980) were probably the first to apply FORM in a comprehensive way for simplified probabilistic design. Their study primarily presented load factors for buildings that were calibrated rationally using FORM and available statistical data. The above review may not do justice to the voluminous research conducted in structural reliability over the past forty or so years. However, it does provide an overview of the historical development of structural LRFD and the accompanying key theoretical advances supporting this development. In the aftermath of recent natural hazards (e.g., Northridge and Kobe earthquakes), the structural engineering profession currently is focusing on performance-based design aimed at meeting client-specific performance goals, in addition to complying with local building codes (Wen 2000, Buckle 2002). An example of a performance criteria matrix proposed by the California Department of Transportation (Caltrans) is shown in Table 1. Efficient techniques for solving time-dependent nonlinear system reliability problems are needed for such problems. Clearly, theoretical developments in structural reliability and applications to probabilistic design are still being pursued actively in the structural community. Geotechnical LRFD The development of geotechnical LRFD has taken a different track. One of the first efforts to rationalize foundation design can be attributed to Hansen (1965), who recommended separate checks for ultimate and serviceability limit states. In contrast, existing WSD often uses the global factor of safety for indirect control of serviceability. Hansen (1965) also recommended the use of partial factors for loads and soil parameters. These partial factors of safety were determined subjectively based on two guidelines: (a) a larger partial factor should be assigned to a more uncertain quantity, and (b) the partial coefficients should result in approximately the same design dimensions as that obtained from traditional practice. Ovesen (1989) highlighted the direct application of a partial factor to the source of uncertainty (soil parameter) as a notable improvement. The partial factors of safety suggested in 1965 were adopted in the Danish Code of Practice for Foundation Engineering (DGI 1978, 1985) with minor modifications. More recent implementations include the Canadian Foundation Engineering Manual, CFEM, third edition (Canadian Geotechnical Society 1992), Geoguide 1, second edition (Geotechnical Engineering Office 1993), and Eurocode 7 (CEN/ TC250 1994). Table 1. Caltrans performance criteria matrix (adapted from Buckle 2002) Bridge Type 1 Bridge Type 2 (Ordinary bridges) (Important bridges) Function evaluation earthquake PL1 PL1 (Frequent earthquake about 200-year DL2 DL1 event) Safety evaluation earthquake PL2 PL1 (Rare earthquake about 1000 to 2000DL3 DL2 year event) where: PL1 - Performance Level 1, characterized by immediate and full access to normal traffic almost immediately after the earthquake PL2 - Performance Level 2, characterized by limited access possible in days, full access within months DL1 - Damage Level 1, characterized by minimal damage with essentially elastic performance DL2 - Damage Level 2, characterized by repairable damage which may executed with minimum loss of functionality DL3 - Damage Level 3, characterized by significant damage which may result in closure but not collapse

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

In North America, the factored resistance approach is the preferred design format. LRFD procedures for bridge superstructures and substructures were introduced in Canada in 1979 as part of the first edition of the Ontario Highway Bridge Design Code, OHBDC1 (Ministry of Transportation Ontario 1979, Green 1991). Green (1991) further noted that these procedures were "basically a simple rearrangement of factor of safety design provisions". The design of deep foundations for power generating stations (Ontario Hydro 1985) broadly followed the second edition, OHBDC2 (Ministry of Transportation Ontario 1983, Klym & Lee 1989). OHBDC is currently in its third edition (OHBDC3), but the foundation resistance factors are not based on reliability calibrations (Green & Becker 2001). In contrast, OHBDC for superstructures was calibrated using reliability theory in its second edition (Grouni & Nowak 1984). The target reliability indices selected were 3.5 for ultimate limit state and 1.0 for serviceability limit state (Nowak & Lind 1979). For fixed offshore platforms, the Canadian standard for foundations, CAN/CSA-S472-92 (CSA 1992a), contains no specification of resistance factors for foundation design, although reliability-based LRFD is available for structural design (Been et al. 1993). Been et al. (1993) further noted that resistance factors were calibrated to the global factor of safety in an earlier 1989 draft commentary (CSA 1989) but were dropped in the 1992 version (CSA 1992b). The main geotechnical design manual in Canada is the Canadian Foundation Engineering Manual, CFEM (Canadian Geotechnical Society 1992). As noted previously, the 1992 version (third edition) is based on the partial factors of safety approach, although the fourth edition currently under preparation is expected to be revised to be consistent with the factored resistance format (Green & Becker 2001). The partial factors of safety in the third edition were calibrated so that they result, on average, in overall factors of safety that are in agreement with existing practice (Meyerhof 1984). To the authors knowledge, CFEM is the only design guide that indirectly recognizes the difficulty of using a single partial factor for each soil parameter to cover the wide range of design equations in which the same soil parameter can appear. Resistance modification factors and performance factors were recommended to ensure more reasonable agreement with existing practice. However, this procedure is not entirely successful, as noted by Baike (1985) and Valsangkar & Schriver (1991). The conflict between the need for simplicity or using small numbers of partial factors of safety, and the need to produce designs comparable with existing practice, does not appear to lend itself readily to simple solutions. The development of LRFD for foundations in the 1995 National Building Code of Canada (NBCC) followed a semi-analytical approach (Becker 1996b). Becker (1996b) opined that a full reliability-based LRFD is difficult to apply because of a lack of statistical data and it is time-consuming and expensive. Therefore, the resistance factors for foundation design were calibrated to fit WSD and to be consistent with a lumped parameter lognormal reliability formula. A target reliability index of 3.5 was used in NBCC for foundation design. As a reference, the NBCC for structural design was calibrated using a target of 3.5 for ductile behavior with normal consequence of failure and a target of 4.0 if either the consequence of failure is severe or the failure mode is brittle (Becker 1996b). In the United States, the resistance factors for design of foundations in the AASHTO LRFD Bridge Code (AASHTO 2002) were derived from NCHRP Report 343 (Barker et al. 1991). The main rationale is to remove the inconsistency between load factor design for superstructures and allowable stress design for foundations, which has resulted in duplication of design efforts because two sets of loads must be evaluated (Rojiani et al. 1991). The resistance factors appear to be determined using a mixture of judgment, calibration with WSD, and reliability analysis. Reliability analysis seems to be used in a limited way (Rojiani et al. 1991, Yoon & O'Neill 1997). The risk levels implied by an extensive range of existing calculation procedures (e.g., rational methods, semi-empirical methods, in-situ methods) formed the basis for the target reliability indices. Paikowsky & Stenersen (2000) further noted that the current AASHTO specifications were developed using insufficient data, hence they utilized mostly backcalculated factors. Most interestingly, the target reliability index for bridge superstructures is 3.5 (Grubb 1997), which is much higher than the target reliability indices quoted in NCHRP Report 343 (2.0 to 2.5, 2.5 to 3.5, and 3.5 for driven piles, drilled shafts and spread footings, respectively). Project NCHRP 2417 was initiated to provide: (a) recommended revisions to the driven pile and drilled shaft portions of section 10 of AASHTO Specifications and (b) detailed procedure for calibrating deep foundation 5

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

resistance factors. The present recommendation is to design for a target reliability of 2.33 if there are five or less piles in a group and 3.00 otherwise (Paikowsky 2002). For API RP 2A-LRFD (API 1993), foundations are treated as one of the structural elements in the RBD calibration process (Moses & Larrabee 1988). A lumped resistance parameter with a bias of 1.0 and a coefficient of variation (COV) of 20% was assumed for pile capacity. The foundation resistance factor was adjusted to achieve an average reliability index of 2.2 for pile axial capacity. A lumped resistance model also was assumed for transmission line structure foundations in the ASCE Manual & Report 74 (Task Committee on Structural Loadings 1991) to preserve a common reliability calibration scheme for structural and foundation components. Resistance factors for a range of lumped resistance COVs (20 to 50%) and target probabilities of failure (0.25% to 1%) were presented. The range of probability of failure corresponds to reliability indices between 2.3 and 2.8. Geotechnical considerations were marginalized because API RP 2A-LRFD and ASCE Manual & Report 74 were focused on structural design. The first attempt to develop simplified RBD specifically for transmission line structure foundations with primary emphasis on geotechnical considerations was described in EPRI Report TR-105000 (Phoon et al. 1995). A number of geotechnical aspects in this study are of general applicability and will be discussed in detail later. RELIABILITY CALIBRATION Basic theory The principal difference between RBD and the traditional or partial factors of safety design approaches lies in the application of reliability theory, which allows uncertainties to be quantified and manipulated consistently in a manner that is free from self-contradiction. A simple application of reliability theory is shown in Figure 1 to define some of the key terms used in RBD. Uncertain design quantities, such as the load (F) and the capacity (Q), are modeled as random variables, while design risk is quantified by the probability of failure (pf). The basic reliability problem is to evaluate pf from some pertinent statistics of F and Q, which typically include the mean (mF or mQ) and the standard deviation (sF or sQ), and possibly the probability density function. A simple closed-form solution for pf is available if both Q and F are normally distributed. For this condition, the safety margin (M = Q - F) also is normally distributed with the following mean (mM) and standard deviation (sM):

mM = mQ mF
s2 M =
2 sQ

(1a)
(1b)

s2 F

Once the probability distribution of M is known, the probability of failure (pf) can be evaluated as:

mM pf = Prob(Q < F) = Prob(Q - F < 0) = Prob(M < 0) = s M

(2)

in which Prob() = probability of an event and () = standard normal cumulative function. Numerical values for () can be obtained easily using the function NORMSDIST(-) in MS Excel. The probability of failure is usually very small for civil infrastructures. A more convenient measure of design risk is the reliability index (), which is defined as: = --1(pf) = mQ mF
2 sQ + s2 F

(3)

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

Figure 1. Reliability assessment for capacity (Q) > load (F) in which -1() = inverse standard normal cumulative function. The function -1 () also can be obtained easily from MS Excel using NORMSINV(). The reliability indices for most geotechnical components and systems lie between 1 and 5, corresponding to probabilities of failure ranging from about 0.16 to 3 10-7, as shown in Figure 2. It is tempting to compare with the traditional factor of safety because both parameters lie in the same range. However, their relationship is actually non-unique, as shown below:

(m

2 FS COVQ ) + COVF 2

m FS 1

(4)

in which mFS = mQ/mF = mean factor of safety and COVQ = sQ/mQ = coefficient of variation (COV) of capacity, COVF = sF/mF = COV of load. Different reliability indices can be obtained for the same mean factor of safety, depending on COVs of Q and F. In this sense, can be considered as an extension and more complete version of FS that attempts to incorporate both deterministic and statistical information on Q and F. The problem of calculating pf for the general case in which Q is modeled as a nonlinear function of several non-normal random variables is more difficult than the simple case shown in Figure 1. A commonly used numerical technique that provides good approximate solutions for engineering applications is the First-Order Reliability Method (FORM). This technique can be easily implemented using MS Excel. Details are given by Low & Phoon (2002).

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

1E+00

Hazardous

Probability of failure

Unsatisfactory Poor 0.16 1E-01 0.07 Below average 0.023 Above average 1E-02 610-3

1E-03 1E-04

10-3 Good 310-5

1E-05
High

1E-06
310-7

1E-07 0 1 2 3 4 5

Reliability index
Figure 2. Relationship between reliability index () and probability of failure (pf) (adapted from Table US Army Corps of Engineers 1997, Table B-1)
Load and resistance factors

The basic objective of RBD is to ensure that the probability of failure of a component does not exceed an acceptable threshold level. Based on this objective, an economical design would be one in which the probability of failure does not depart significantly from the threshold. For the design problem shown in Figure 1, the RBD objective can be formally stated as follows: pf = Prob(Q < F) pT (5) in which pT = acceptable target probability of failure. Hence, the only basic difference between RBD and existing practice is that one controls the probability of failure, rather than the factor of safety. In nonmathematical terms, this is equivalent to controlling the fraction of unacceptable solutions in a weighted parametric study, where the weights refer to the likelihood of a set of input parameters being correct. Reliability-based design, as exemplified by Equation 5, allows the engineer to make a conscious choice on an acceptable level of design risk and then proceed to a set of design dimensions that are consistent with that choice. In contrast to the traditional or partial factors of safety approach, logical consistency between the computed design risk and the uncertainties inherent in the design process is assured by probabilistic analysis, such as FORM. A log-normal probability model is commonly used in place of the simple normal probability model because most physical quantities are non-negative (e.g., Paikowsky 2002). The analytical solution is available but the following well-known approximation is typically used for reliability calibration (e.g., Ravindra & Galambos 1978, Becker 1996a):

log e (m FS )
2 2 COVF + COVQ

(6)

To derive the LRFD format, the denominator in Equation 6 must be linearized as follows (Lind 1971):
2 2 COVF + COVQ 0.75(COVF + COVQ )

(7)

Using Equations 6 and 7, the following simple LRFD format can be obtained:

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

Qn = Fn in which Qn, Fn = nominal capacity and load, and , = resistance and load factors given by: = mQ Qn exp( 0.75 COVQ )

(8)

(9a)

mF exp(+ 0.75 COVF ) Fn

(9b)

It can be seen that reliability-calibrated resistance and load factors include the target risk level () and the underlying parametric uncertainties (COVQ and COVF) rationally into the design process. Reservations have been expressed about the increase in the complexity of the design calculations resulting from the use of reliability theory. Complicated reliability calculations are undesirable because: (a) statistical information is not sufficiently well-defined to warrant sophisticated treatment, (b) there is a greater risk of making computational errors, (c) a study in soil behavior prediction is reduced to a mere mathematical exercise, and (d) attention will be diverted from the proper characterization of the ground mass and appreciation of the physical, chemical, and mechanical processes taking place in it (Beal, 1979; Semple, 1981; Simpson, et al., 1981; Boden, 1981). These reservations are not without merits. Excessive preoccupation with maintaining simplicity would, however, ultimately be a disservice to the geotechnical engineering profession. Historical hindsight has shown clearly that the judicious use of rational methods, as initiated by Terzaghi in 1943, was the primary cause of most of the significant advances in soil mechanics following World War II. The cost to pay for rationality is that design calculations could become more complicated. However, this cost is more than offset by the benefits associated with the use of rational methods. For example, the improvement in soil behavior prediction allows less conservatism to be applied in the design. The use of reliability methods is the next logical step toward greater rationality in design, and their potential benefits should not be discarded heedlessly because of the reluctance to advance beyond the current level of complexity in design. GEOTECHNICAL UNCERTAINTIES
Parametric uncertainty

The evaluation of soil and rock properties is one of the key design aspects that distinguishes geotechnical from structural engineering. None of the current geotechnical LRFD implementations consider this important issue explicitly. The purpose of this section is to highlight two important observations: (a) geotechnical variability is a complex attribute that needs careful evaluation, and (b) extensive statistical data are available for use as first-order estimates in RBD calibration and application. Extensive calibration studies by Phoon et al. (1995) indicated that foundation resistance factors in the RBD equations can be calibrated for broad categories of data quality (e.g., COV of undrained shear strength = 10-30%, 30-50%, 50-70%) without compromising on the uniformity of reliability achieved. Hence, it is not true that there are insufficient statistical data to warrant realistic reliability calculations. There are three primary sources of geotechnical uncertainties: (a) inherent variabilities, (b) measurement uncertainties, and (c) transformation uncertainties. The first results primarily from the natural geologic processes that produced and continually modify the soil mass in-situ. The second is caused by equipment, procedural and/or operator, and random testing effects. Equipment effects result from inaccuracies in the measuring devices and variations in equipment geometries and systems employed for routine testing. Procedural and/or operator effects originate from the limitations in existing test standards and how they are followed. In general, tests that are highly operator-dependent and have complicated test procedures will have greater variability than those with simple procedures and little operator dependency, as described in detail elsewhere (Kulhawy & Trautmann 1996). Random testing
9

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

error refers to the remaining scatter in the test results that is not assignable to specific testing parameters and is not caused by inherent soil variability. The third component of uncertainty is introduced when field or laboratory measurements are transformed into design soil properties using empirical or other correlation models (e.g., correlating the standard penetration test N value with the undrained shear strength). Obviously, the relative contribution of these components to the overall uncertainty in the design soil property depends on the site conditions, degree of equipment and procedural control, and quality of the correlation model. Therefore, soil property statistics that are determined from total variability analyses only can be applied to the specific set of circumstances (site conditions, measurement techniques, correlation models) for which the design soil properties were derived. Useful guidelines on typical coefficients of variation of many common soil properties are summarized in Tables 2 to 5. Others are reported by Jones et al. (2002).
Model uncertainty

A similar effort is underway to quantify uncertainties associated with geotechnical calculation models. Although many geotechnical calculation models are simple, reasonable predictions of fairly complex soil-structure interaction behavior still can be achieved through empirical calibrations. Because of our geotechnical heritage that is steeped in such empiricisms, model uncertainties can be significant. Even a simple estimate of the average model bias is crucial for reliability-based design. If the model is conservative, it is obvious that the probabilities of failure calculated subsequently will be biased, because those design situations that belong to the safe domain could be assigned incorrectly to the failure domain, as a result of the built-in conservatism. Robust model statistics can only be evaluated using: (a) realistically large-scale prototype tests, (b) a sufficiently large and representative database, and (c) reasonably high quality testing where extraneous uncertainties are well-controlled. With the possible exception of foundations, insufficient test data are available to perform robust statistical assessment of the model error in many geotechnical calculation models. The development of a fully rigorous reliability-based design code that can handle the entire range of geotechnical design problems is currently impeded by the scarcity of these important statistics. Sidi (1986) was among the first to report model statistics that were established firmly using a large load test database assembled by Olson & Dennis (1982). The focus of the study was on friction piles in clay subjected to axial loading. Briaud & Tucker (1988) conducted a similar study using a 98-pile load test database obtained from the Mississippi State Highway Department. Recent literature includes estimation of model statistics for the calibration of deep foundation resistance factors for AASHTO [American Association of State Highway and Transportation Officials] (Paikowsky 2002). A substantial part of the study pertains to the evaluation of driven pile axial capacity using dynamic methods. None of these studies addresses the applicability of model statistics beyond the conditions implied in the database. This question mirrors the same concern expressed previously on the possible site-specific nature of soil variabilities. Phoon & Kulhawy (2003) presents a critical evaluation of model factors using an extensive database collected as part of an EPRI (Electric Power Research Institute) research program on transmission line structure foundations (Chen & Kulhawy 1994). This study considers only free-head rigid drilled shafts, using the databases summarized in Table 6. Three common models for lateral soil resistance, by Reese, Hansen, and Broms, are used to compute the theoretical lateral capacity under undrained and drained loading modes. Details are given elsewhere (Chen & Kulhawy 1994). Also, there are several methods to interpret lateral capacity from load tests: displacement limit, rotation limit, lateral or moment limit, and hyperbolic capacity (Hirany & Kulhawy 1988). All theoretical models are based on failure mechanisms requiring full mobilization of soil strength. Therefore, the hyperbolic capacity is probably closest to this ultimate state assumed in the analyses. A plausible and common method of correcting for model error is to assume the following multiplicative model (e.g., Ang & Tang 1984, Sidi 1986): Hh = M Hu (10)

10

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

Table 2. Approximate guidelines for inherent soil variability (Source: Phoon and Kulhawy 1999a). Soil type Mean COV(%) Test type Propertya Lab strength su(UC) su(UU) su(CIUC)
'

Clay Clay Clay Clay & sand Clay Clay Sand Clay Clay & sand Clay Sand Clay Sand Sand Sand Sand Clay Sand Sand Clay & silt Clay & silt Clay & silt Clay & silt Clay & silt Clay & silt Sand

10-400 kN/m 2 10-350 kN/m 2 150-700 kN/m o 20-40 0.5-2.5 MN/m 2 0.5-2.0 MN/m 2 0.5-30.0 MN/m 5-400 kN/m
2 2

20-55 10-30 20-40 5-15 < 20 20-40 20-60 10-40 25-50 10-35 20-50 10-35 20-50 20-60 20-60 15-65 10-35 20-50 15-65 8-30 6-30 6-30 b b < 10 c 10-40 d 50-70

CPT

qT qc su(VST) N A reading B reading ID KD ED

VST SPT DMT

10-70 blows/ft 100-450 kN/m 2 60-1300 kN/m 2 500-880 kN/m 2 350-2400 kN/m 1-8 2-30 2 10-50 MN/m 400-2800 kN/m 2 1600-3500 kN/m 2 5-15 MN/m 13-100 % 30-90 % 15-25 % 10-40 % 10 % 3 13-20 kN/m 30-70 %
2 2

PMT

pL EPMT

Lab index

wn wL wP PI LI , d Dr

a - su = undrained shear strength; UC = unconfined compression test; UU = unconsolidated-undrained triaxial compression test; CIUC = consolidated isotropic undrained triaxial compression test; ' = effective stress friction angle; qT = corrected cone tip resistance; qc = cone tip resistance; VST = vane shear test; N = standard penetration test blow count; A & B readings, ID, KD, & ED = dilatometer A & B readings, material index, horizontal stress index, & modulus; pL & EPMT = pressuremeter limit stress & modulus; wn = natural water content; wL = liquid limit; wP = plastic limit; PI = plasticity index; LI = liquidity index; & d = total & dry unit weights; Dr = relative density b - COV = (3-12%) / mean c - total variability for direct method of determination d - total variability for indirect determination using SPT values

in which Hh = measured hyperbolic capacity, Hu = computed lateral capacity, and M = model factor, typically assumed to be a log-normal random variable. The empirical distributions of M for model-scale laboratory tests and full-scale field tests are summarized in Figure 3.

11

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

Table 3. Summary of total measurement error of some laboratory tests (Phoon & Kulhawy 1999a) No. No. Tests / Property Value Property data Group (unitsb) COV (%) Soil type groups Range Mean Range Mean Range Mean Propertya su(TC) su(DS) su(LV) Clay, silt Clay, silt Clay Clay, silt Clay, silt Sand tan (TC) Sand, silt tan (DS) Clay wn wL wP PI Fine-grained Fine-grained Fine-grained Fine-grained Fine-grained 11 2 15 4 5 2 6 2 3 26 26 10 3 13-17 9-13 9-13 26 82-88 41-89 41-89 41-89 82-88 13 15 10 11 26 85 64 62 61 85 7-407 108-130 4-123 2-27 24-40
o

125 119 29 19.1 33.3 32.7 18 36 21 23 17.0


o

8-38 19-20 5-37 7-56 3-29 13-14 2-22 6-22 6-12 3-11 7-18 5-51 1-2

19 20 13 24 13 14 8 14 8 7 10 24 1

(TC) (DS)

o o

o o

30-35 -

16-21 17-113 12-35 4-44 16-17

a - su = undrained shear strength; = effective stress friction angle; TC = triaxial compression test; UC = unconfined compression test; DS = direct shear test; LV = laboratory vane shear test; wn = natural water content; wL = liquid limit; wP = plastic limit; PI = plasticity index; = total unit weight 2 3 b - units of su = kN/m ; units of wn, wL, wP, and PI = %; units of = kN/m

The model-scale load tests were conducted in uniform kaolinite clay and filter sand deposits prepared under controlled laboratory conditions. Hence, uncertainties arising from evaluation of soil parameters are minimal. In addition, construction variabilities and measurement errors associated with load tests also are minimal. Therefore, model uncertainties computed from laboratory tests should be an accurate indicator of errors arising from the use of simplified calculation models. The main concern is whether the model factors are applicable beyond the uniform profile and specific soil type used in the laboratory. Model factors from field tests are expected to be more general because they are computed from load tests conducted in more diverse site environments. However, it is reasonable to query if the statistics of such model factors are lumped statistics, in the sense that extraneous sources of uncertainties (e.g., construction variabilities, measurement errors incurred during load test, uncertainties in soil parameter evaluation) are inextricably included in the computation.

12

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

Table 4. Summary of measurement error of common in-situ tests (Kulhawy & Trautmann 1996) Coefficient of variation, COV (%) Test Equipment Procedure Random Totala Rangeb Standard penetration test (SPT) Mechanical cone penetration test (MCPT) Electric cone penetration test (ECPT) Vane shear test (VST) Dilatometer test (DMT) Pressuremeter test, pre-bored (PMT) Self-boring pressuremeter test (SBPMT)
a b
2

5c - 75d 5 3 5 5 5 8
2

5c - 75d 10e -15f 5 8 5 12 15

12 - 15 10e -15f 5e -10f 10 8 10 8


2 0.5

14c - 100d 15e -22f 7e - 12f 14 11 16 19

15 - 45 15 - 25 5 - 15 10 - 20 5 - 15 10 - 20g 15 - 25g

- COV(Total) = [COV(Equipment) + COV(Procedure) + COV(Random) ] - Because of limited data and judgment involved in estimating COVs, ranges represent probable magnitudes of field test measurement error c, d - Best to worst case scenarios, respectively, for SPT e, f - Tip and side resistances, respectively, for CPT g - It is likely that results may differ for po, pf, and pL, but the data are insufficient to clarify this issue

A comparison between laboratory and field data such as that shown in Figure 3 is illuminating. Laboratory and field results are plotted as white and grey histograms in the first row of each cell, respectively. Visual inspection and simple statistics [mean, standard deviation (S.D.), coefficient of variation (COV)] show that the histograms are similar. The p-values from the Mann-Whitney test formally show that the null hypothesis of equal medians cannot be rejected at 5% significance level, with the exception of the drained factor for the Hansen model. Therefore, it is reasonable to argue that the results presented in Figure 3 have wider applicability beyond the conditions implied by the underlying databases, and the model uncertainties are mainly caused by errors intrinsic to the respective simplified calculation models. A more robust estimate of the empirical distribution is obtained by combining the laboratory and field data, as shown in the second row of each cell in Figure 3. For rigid drilled shafts subjected to lateral-moment loading, the COV of the model factor appears to fall within a narrow range of 25 to 40%. Another detailed example on cantilever walls in sand is discussed by Phoon et al. (2003d). PRACTICAL IMPLEMENTATION ISSUES
Conceptual basis for EPRI study

Although existing geotechnical LRFD codes look the same as their structural counterparts, they are incompatible with structural RBD on closer inspection because one or more of the following key elements are missing (Kulhawy & Phoon 2002): a. The primary objective in structural RBD is to achieve a minimum target reliability index across a specified domain of interest (e.g., foundation geometries and types, loading modes, soil conditions, etc.). Structural RBD requires deliberate and explicit choices to be made on the target reliability 13

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

Table 5. Approximate guidelines for design soil property variability (Phoon & Kulhawy 1999b) Design Point Spatial avg. Correlation a b c Soil type COV (%) equation property Test COV (%) su(UC) su(UU) su(CIUC) su(field) su(UU) su(CIUC) su(UU) d su su(field)
(TC)

Direct (lab) Direct (lab) Direct (lab) VST qT qT N KD PI Direct (lab) qT PI Direct (SBPMT) Direct (SBPMT) KD N Direct (PMT) Direct (DMT) N N

Clay Clay Clay Clay Clay Clay Clay Clay Clay Clay, sand Sand Clay Clay Sand Clay Clay Sand Sand Clay Silt

20-55 10-35 20-45 15-50 e 30-40 e 35-50 40-60 30-55 e 30-55 7-20 10-15 15-20
e e

10-40 7-25 10-30 15-50 e 30-35 e 35-40 40-55 30-55 6-20 10


e e

14 18 18 23 29 32 38 43 49 54 61 64

cv

15-20

Ko Ko Ko Ko EPMT ED EPMT ED

20-45 25-55 e 35-50 e 40-75 20-70 15-70 85-95 40-60

15-45 20-55 e 35-50 15-70 10-70 85-95 35-55

a - su = undrained shear strength; UU = unconsolidated-undrained triaxial compression test; UC = unconfined compression test; CIUC = consolidated isotropic undrained triaxial compression test; su(field) = corrected su from vane shear test; = effective stress friction angle; TC = triaxial compression; cv = constant volume ; Ko = in-situ horizontal stress coefficient; EPMT = pressuremeter modulus; ED = dilatometer modulus b - VST = vane shear test; qT = corrected cone tip resistance; N = standard penetration test blow count; KD = dilatometer horizontal stress index; PI = plasticity index c - averaging over 5 meters d - mixture of su from UU, UC, and VST e - COV is a function of the mean; refer to COV equations in Phoon & Kulhawy (1999b) for details

Table 6. Databases on laterally-loaded rigid drilled shafts (Phoon & Kulhawy 2003) Description No. tests D/B Undrained loading: Model-scale lab tests 47 3.0 8.0 Full-scale field tests 27 2.3 10.5 Drained loading: Model-scale lab tests 55 2.6 9.0 Full-scale field tests 22 2.5 7.0
Note: D = depth; B = diameter; e = load eccentricity

e/D 0.03 4.0 0.03 6.8 0.06 5.0 0 5.4

14

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

Frequency

Drained: Reese et al. (1974)

10 5 0 0.4 20

Frequency

Model Undrained: Reese (1958)

Undrained
20 15
Mean = S.D. = COV = n= p= Lab 1.43 0.37 0.26 47 0.315 Field 1.40 0.47 0.33 27

Drained
20 15 10 5 0
Mean = S.D. = COV = n= p= Lab 1.19 0.57 0.48 55 0.433 Field 1.19 0.36 0.30 22

1.2 2.0 2.8 Hh/H u(Reese)


Mean = S.D. = COV = n= 1.42 0.41 0.29 74

3.6 20 Frequency 15 10 5 0

0.4

1.2 2.0 2.8 Hh/H u(Reese)


Mean = S.D. = COV = n= 1.19 0.51 0.43 77

3.6

Frequency

15 10 5 0 0.4

1.2 2.0 2.8 H h/Hu(Reese)


Mean = S.D. = COV = n= p= Lab 1.95 0.55 0.28 47 0.296

3.6
Field 1.85 0.57 0.31 27

0.4 20 Frequency 15 10 5 0

1.2 2.0 2.8 H h/Hu(Reese)


Mean = S.D. = COV = n= p= Lab 1.05 0.33 0.31 55 0.002

3.6
Field 0.83 0.25 0.30 22

Hansen (1961)
Frequency

20 15 10 5 0 0.4 20 Frequency 15 10 5 0 0.4

1.2 2.0 2.8 Hh/H u(Hansen)


Mean = S.D. = COV = n= 1.92 0.56 0.29 74

3.6
20 Frequency 15 10 5 0

0.4

1.2 2.0 2.8 H h/H u(Hansen)

3.6

Mean = S.D. = COV = n=

0.98 0.32 0.33 77

1.2 2.0 2.8 H h/H u(Hansen)


Lab 2.28 0.80 0.35 47 0.875

3.6

0.4

1.2 2.0 2.8 H h/Hu(Hansen)


Mean = S.D. = COV = n= p= Lab 1.31 0.53 0.40 55 0.844

3.6
Field 1.27 0.41 0.32 22

Drained: Broms (1964b)

10 5 0 0.4 20

Frequency

Frequency

Undrained: Broms (1964a)

20 15
Mean = S.D. = COV = n= p=

Field 2.29 0.95 0.41 27

20 15 10 5 0

1.2 2.0 2.8 H h/Hu(Broms)


Mean = S.D. = COV = n= 2.28 0.85 0.37 74

3.6 20 Frequency 15 10 5 0

0.4 1.2 2.0 2.8 3.6 H h/H u(simplified Broms)


Mean = S.D. = COV = n= 1.30 0.50 0.38 77

Frequency

15 10 5 0 0.4

1.2 2.0 2.8 Hh/H u(Broms)

3.6

0.4 1.2 2.0 2.8 3.6 H h/Hu(simplified Broms)

Figure 3. Model factors for rigid drilled shafts under undrained and drained lateral-moment loading modes (Phoon & Kulhawy 2003)

15

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

index, scope of calibration domains, and representative designs populating each domain. This is philosophically different from the objective of achieving designs comparable to working stress design. b. The secondary objective in structural RBD is to increase uniformity of reliability across the domain of interest, which is rarely emphasized and verified in geotechnical LRFD. In fact, the typical use of a single resistance factor for each loading mode is not adequate for this task, as elaborated in item 6 below. c. Soil variability is the most significant source of uncertainty, but it is not quantified in a robust way (if at all) and incorporated explicitly in the code calibration process. d. Probabilistic load models compatible with the relevant structural codes are not spelled out clearly. It is unclear if the original structural load models have been used for code calibration. Load combinations are not amenable to simplified lognormal reliability analysis (in contrast to the more general first-order reliability method or FORM highlighted in item 5) unless they are approximated as some lumped load parameters. e. Rigorous reliability analysis using FORM is not used as the main tool to integrate loads, basic soil parameters, and calculation models in a realistic and self-consistent way, both physically and probabilistically. The commonly adopted approach of simplifying geotechnical capacity as a single lognormal random variable has limitations (Phoon et al. 2003b). f. No guidelines on selection of nominal or characteristic soil parameters are usually given. It is also unclear how resistance factors will be affected by the site conditions, measurement techniques, and correlation models used to derive the relevant design parameters. There are no technical and/or practical difficulties in addressing these incompatibilities directly. The basic solution is to follow a more general calibration procedure as outlined below.
Achieving uniform reliability

In EPRI Report TR-105000, uniform reliability is realized by partitioning the design domains and using a Multiple Resistance Factor Design (MRFD) format. The design of drilled shafts (bored piles) for uplift under undrained loading will be used as an example. Two simple design formats were selected for reliability calibration: LRFD: F50 uQun MRFD: F50 suQsun + tuQtun + wW (11) (12)

in which F50 = 50-year return period load, Qun = nominal uplift capacity, Qsun = nominal side resistance, Qtun = nominal tip resistance, W = weight of foundation, and u, su, tu and w = resistance factors. Figure 4a show the reliability levels implicit in existing ULS designs. Note that the existing WSD format is essentially the same as the LRFD format (Equation 11), because the reciprocal of the traditional factor of safety (FS) is equal to the resistance factor (u). Therefore, the variation in the reliability index () or probability of failure (pf) at a fixed FS is indicative that the LRFD format will produce fairly poor uniformity in reliability when it is applied over the entire design domain. In EPRI Report TR-105000, the uniformity in reliability was improved by using the following general calibration procedure: a. Perform a parametric study on the variation of the reliability level with respect to each deterministic and statistical parameter in the design problem. Examples of deterministic parameters that control the design of drilled shafts include the diameter (B) and depth to diameter (D/B) ratio. Examples of statistical parameters include the mean and coefficient of variation (COV) of the undrained shear strength (su). b. Partition the parameter space into several smaller domains. An example of a simple parameter space is shown in Figure 5. The reason for partitioning is to achieve greater uniformity in reliability over the full range of deterministic and statistical parameters. For those parameters identified in Step (a) as having a significant influence on the reliability level, the size of the partition clearly should be smaller. In addition, partitioning ideally should conform to existing geotechnical conventions. 16

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

(b) (a) Figure 4. Drilled shafts in undrained uplift: (a) reliability levels implicit in existing ultimate limit state design and (b) performance of ultimate limit state RBD formats c. Select a set of representative points from each domain. Note that each point in the parameter space denotes a specific set of parameter values (Figure 5). Ideally, the set of representative points should capture the full range of variation in the reliability level over the whole domain. d. Determine an acceptable foundation design for each point and evaluate the reliability levels in the designs. Foundation design is performed using the set of parameter values associated with each point, along with a simplified RBD format and a set of trial resistance factors. The reliability of the resulting foundation design then is evaluated using the FORM algorithm. e. Quantify the deviations of the reliability levels from a pre-selected target reliability index, T. The following simple objective function can be used:

H(su , tu , w ) =

(
i =1

- T ) 2

(13)

f.

in which H() = objective function to be minimized, n = number of points in the calibration domain, and i = reliability index for the ith point in the domain. Adjust the resistance factors and repeat Steps (d) and (e) until the objective function is minimized. The set of resistance factors that minimizes the objective function (H) is the most desirable because the degree of uniformity in the reliability levels of all the designs in the domain is maximized. The following measure can be used to quantify the degree of uniformity that has been achieved:
= H n

(11)

g. Repeat Steps (c) to (f) for the other domains.


17

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

Figure 5. Partitioning of parameter space for calibration of resistance factors. Table 7. Undrained uplift resistance/deformation factors for drilled shafts (Source: Phoon et al. 1995, pp. 6-7 & 17). LRFD MRFD LRFD COV su Mean su (ULS) (ULS) (SLS) (%) (kN/m2) su tu w u u 25 - 50 10 30 0.44 0.44 0.28 0.50 0.65 Medium clay 30 50 0.43 0.41 0.31 0.52 0.63 50 70 0.42 0.38 0.33 0.53 0.62 50 - 100 Stiff clay 100 - 200 Very stiff clay 10 30 30 50 50 70 10 30 30 50 50 70 0.43 0.41 0.39 0.40 0.37 0.34 0.40 0.36 0.32 0.35 0.31 0.26 0.35 0.37 0.40 0.42 0.48 0.51 0.56 0.59 0.62 0.66 0.68 0.72 0.64 0.61 0.58 0.61 0.57 0.52

Note: Target = 3.2 for ultimate limit state (ULS) and 2.6 for serviceability limit state (SLS)

The results of the RBD calibration exercise for drilled shafts in undrained uplift loading are shown in Table 7. Exact comparison with other LRFD resistance factors is difficult, but AASHTO recommends u = 0.55 for uplift capacity of drilled shaft (-method) in clay (NCHRP Report 343, Table 4.10.6-3) while OHBDC/CHBDC and NBCC recommends u = 0.3 for tension capacity of deep foundations evaluated using static analysis (Green & Becker 2001). Note that the target reliability indices are 2.5 to 3.5 for AASHTO (probably closer to 2.5 as reported by Rojiani et al. 1991) and 3.5 for OHBDC/CHBDC/NBCC. The EPRI resistance factors vary from 0.34 to 0.44 depending on the quality of data, as shown in Table 7. They lie between AASHTO and OHBDC/CHBDC/NBCC factors, but are closer to the latter than the former, partially because the target reliability index is 3.2. More importantly, a single resistance factor cannot be expected to maintain uniform reliability over a wide and diverse range of design scenarios as shown in Figure 4a. The EPRI study shows that the use of a simple 3x3 partitioning on the mean and COV of undrained shear strength is sufficient to produce designs with distinctively more uniform 18

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

reliability (compare Figures 4a and 4b). The partitioning on the mean undrained shear strength also was selected to conform to existing geotechnical conventions, as noted previously. The EPRI study further recommended use of the Multiple Resistance Factor Design (MRFD) format for achieving a more consistent target reliability (Equation 12). The MRFD format is a natural generalization of the LRFD format that involves the application of one resistance factor to each component of the capacity rather than the overall capacity. MRFD is more physically meaningful for foundation design because the variability of each component can be significantly different. In addition, it achieves greater uniformity in reliability as shown in Figure 4b. CONCLUSIONS New reliability-based design (RBD) methodologies that are already adopted widely by the structural community are not accepted readily in the geotechnical community, partially because of the questionable robustness of the statistics used for code calibration and unfamiliarity with probabilistic concepts. However, maintaining status quo is increasingly untenable because of gathering momentum in code harmonization and broadening divergence between geotechnical and structural design. Because geotechnical design is only one component of harmonised codes, it is anticipated that structural reliability methods will eventually prevail in geotechnical design. There is a need to draw a clear distinction between accepting reliability analysis as a necessary theoretical basis for geotechnical design and downstream calibration of simplified multiple-factor design formats, with emphasis on the former. Simplified reliability-based design (RBD) equations in the form of LRFD/MRFD are probably required for routine design at present, but their limitations have no bearing on the generality of reliability theory. This paper argues that there is sufficient statistical support for the development of these simplified RBD formats. Reliability-based design, simplified or otherwise, provides a more consistent means of managing uncertainties, but it is by no means a perfect solution. Engineering judgment still is indispensable in many aspects of geotechnical engineering reliability analysis merely removes the need for guesswork on how uncertainties affect performance and is comparable to the use of elasto-plastic theory to remove the guesswork on how loads induce stresses and deformations. Lacasse et al. (2004) observed that: Engineering depends on judgment, the exercise of which depends on knowledge derived from theoretical concepts, experiment, measurements, observations, and past experience. These building blocks have to be recognized, assembled, and evaluated collectively before judgment can be rendered. Historical hindsight has shown clearly that the judicious use of rational methods, as initiated by Terzaghi in 1943, was the primary cause of most of the significant advances in soil mechanics following World War II. The cost to pay for rationality is that design calculations could become more complicated. However, this cost is more than offset by benefits such as reduction in conservatism because of improved understanding and hence increased confidence in the design. The use of reliability methods is the next logical step toward greater rationality in design, and their potential benefits should not be discarded heedlessly because of the reluctance to advance beyond the current level of complexity in design. ACKNOWLEDGMENTS Financial support from the Site Investigation Committee, Korean Geotechnical Society; Hyundai Engineering & Construction Co., Ltd.; and Daelim Industry Co., Ltd are gratefully acknowledged. The author is also grateful to Dr Gil Lim Yoon for his kind invitation to deliver this special lecture and making all the arrangements for this visit.

19

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

REFERENCES Ang, A. H-S. & Tang, W. H. 1984. Probability concepts in engineering planning and design (2). John Wiley & Sons, New York. American Association of State Highway and Transportation Officials (AASHTO) 2002. LRFD Highway Bridge Design Specifications, Second Edition, AASHTO, Washington. American Petroleum Institute (API) 1993. Recommended practice for planning, designing and constructing fixed offshore platforms - Load and Resistance Factor Design, API RP 2A-LRFD, Washington. Baike, L. D. 1985. Total and partial factors of safety in geotechnical engineering. Can. Geotech. J., 22(4): 477-482. Barker, R M, Duncan, J. M., Rojiani, K. B., Ooi, P. S. K., Tan, C. K. & Kim, S. G. 1991. Manuals for design of bridge foundations, NCHRP Report 343, Transportation Research Board, Washington. Beal, A. N. 1979. What's wrong with Load Factor Design? In Proc. Institution of Civil Engineers, 66(Pt 1): 595 - 604. Been, K., Clark, J. I. & Livingstone, W. R. 1993. Verification and calibration studies for the new CAN/CSA-S472 foundations of offshore structures. Can. Geotech J., 30(3): 515-525. Becker, D E. 1996a. Limit states design for foundations. Part I. An overview of the foundation design process. Can. Geotech. J., 33(6): 956-983. Becker, D E. 1996b. Limit states design for foundations. Part II. Development for National Building Code of Canada. Can. Geotech. J. 33(6): 984-1007. Boden, B. 1981. Limit state principles in geotechnics. Ground Engrg., 14(6): 2 - 7. Bolton, M. D. 1989. Development of codes of practice for design. In Proc. 12th Int. Conf. Soil Mech. Fdn. Engrg., (3): 2073 2076, Rio de Janeiro. Briaud, J.L. & Tucker, L.M. 1988. Measured and predicted axial response of 98 piles. J. Geotech. Engrg., ASCE, 114(9): 984-1001. Broms, B. B. 1964a. Lateral resistance of piles in cohesive soils. J. Soil Mech. Fdn. Div., ASCE, 90(SM2): 27-63. Broms, B. B. 1964b. Lateral resistance of piles in cohesionless soils. J. Soil Mech. Fdn. Div., ASCE, 90(SM3): 123-156. Buckle, I. G. 2002. AASHTO LRFD limit state design of bridges with emphasis on seismic performance. In Proc. International Workshop on Foundation Design Codes and Soil Investigation in view of International Harmonization and Performance Based Design, 17-30, Tokyo. Canadian Geotechnical Society 1992. Canadian Foundation Engineering Manual (CFEM), 3rd Ed., Richmond. Canadian Standards Association (CSA) 1989. Commentary to CSA preliminary standard S472-M1989, Foundations, Ont. Publ. No. S472.1-M1989, Rexdale, Ontario. Canadian Standards Association (CSA) 1992a. Foundations, Ont. Publ. No. CAN/CSA-S472-92, Rexdale, Ontario. Canadian Standards Association (CSA) 1992b. Commentary to CSA standard CAN/CSA-S472-92, Foundations, Ont. Publ. No. S472.1-92, Rexdale, Ontario. Canadian Standards Association (CSA) 2000. Canadian Highway Bridge Design Code (CHBDC), CSA Standard S6-00, Rexdale, Ontario. CEN/TC250 1994. Geotechnical design - Part 1, General rules, Eurocode 7, ENV-1997-1, European Committee for Standardization (CEN). Chen, Y-J. & Kulhawy, F. H. 1994. Case history evaluation of behavior of drilled shafts under axial and lateral loading. Report TR-104601, Electric Power Research Institute, Palo Alto. Committee on Reliability Methods for Risk Mitigation in Geotechnical Engineering 1995. Probabilistic Methods in Geotechnical Engineering, National Academy of Sciences, Washington. Cornell, C. A. 1969. A probability-based structural code. J. Amer. Conc. Inst., 66(12): 974-985. Danish Geotechnical Institute (DGI) 1978. Code of practice for foundation engineering, Bulletin 32, Copenhagen. 20

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

Danish Geotechnical Institute (DGI) 1985. Code of practice for foundation engineering, Bulletin 36, Copenhagen. DiMaggio, J., Saad, T., Allen T., Christopher, B. R., Dimillio, A., Goble, G., Passe P., Shike, T. & Person, G. 1999. Geotechnical engineering practices in Canada and Europe. Report FHWA-PL-99-O, Federal Highway Administration (FHWA), Washington. Duncan, J. M. 2000. Factors of safety and reliability in geotechnical engineering. J. Geotech. Geoenviron. Engrg., ASCE, 126(4): 307-316. Ellingwood, B. R., Galambos, T. V., MacGregor, J. G. & Cornell, C. A. 1980. Development of probability-based load criterion for American National Standard A58. Special Publication 577, National Bureau of Standards, Washington. Frank, R. 2002. Development and future of Eurocode 7 Geotechnical design. In Proc. Int. Workshop on Foundation Design Codes and Soil Investigation in view of International Harmonization and Performance Based Design, 51 58, Tokyo. Freudenthal, A. M. 1947, Safety of Structures. Transactions, ASCE, 112:125-159. Geotechnical Engineering Office 1993. Guide to retaining wall design. Geoguide 1, 2nd Ed., Hong Kong. Goble, G. 1999. Geotechnical related development and implementation of load and resistance factor design (LRFD) methods. NCHRP Synthesis 276, Transportation Research Board, Washington. Golder, H. Q. 1966. Discussion of "Role of the calculated risk in earthwork and foundation engineering". J. Soil Mech. Fdn. Div., ASCE, 92(SM1): 188-189. Green, R. 1991. The development of a LRFD code for Ontario bridge foundations. In Proc. Geotech. Engrg. Congress (GSP27), ASCE, 1365-1376, New York. Green, R. & Becker, D. 2001. National report on limit state design in geotechnical engineering: Canada. Geotechnical News, 19(3): 47-55. Grouni, H. N. & Nowak, A. S. 1984. Calibration of the Ontario Bridge Design Code 1983 edition. Can. J. Civ. Engrg., 11: 760-770. Grubb, M. A. 1997. LFD vs. LRFD - What's up with the letter "R" anyway? Bridge Crossings 5: 1-3. Download from www.aisc.org/msc/BridgeXings_No05.pdf. Hansen, J. B. 1961. Ultimate resistance of rigid piles against transversal forces. Bulletin 12, Danish Geotechnical Institute, 5-9, Copenhagen. Hansen , J. B. 1965. Philosophy of foundation design: Design criteria, safety factors and settlement limits. In Proc. Symp. on Bearing Capacity & Settlement of Foundations: 1-13, Duke University, Durham. Hasofer, A. M. & Lind, N. C. 1974. Exact and invariant second-moment code format. J. Engrg. Mech. Div., ASCE, 100(EM1): 111-121. Hirany, A. & Kulhawy, F. H. 1988. Conduct and interpretation of load tests on drilled shaft foundation: detailed guidelines. Report EL-5915 (1), Electric Power Research Institute, Palo Alto. Jones, A.L., Kramer, S.L. & Arduino, P. 2002. Estimation of uncertainty in geotechnical properties for performance-based earthquake engineering. PEER Report 2002/16, Pacific Earthquake Engineering Research Center, University of California, Berkeley. Klym, T. W. & Lee, C. F. 1989. Limit states foundation design practice at Ontario Hydro. In Proc. Symp. Limit State Design in Fdn. Engrg., 161 - 169, Toronto. Kulhawy, F. H. 1984. ASCE drilled shaft standard: University perspective. Analysis & Design of Pile Foundations, ASCE: 390-395, New York. Kulhawy, F. H. 1996. From Casagrande's "Calculated Risk" to reliability-based design in foundation engineering. Civil Engrg. Prac., BSCE, 1(2): 43-56. Kulhawy, F. H. & Phoon, K. K. 1996. Engineering judgment in the evolution from deterministic to reliability-based foundation design. In Uncertainty in the Geologic Environment - From Theory to Practice (GSP 58), ASCE: 29-48, New York. Kulhawy, F. H. & Phoon, K. K. 2002. Observations on geotechnical reliability-based design development in North America. In Proc. Int. Workshop on Foundation Design Codes & Soil Investigation in view of International Harmonization & Performance Based Design, 31-48, Tokyo. Kulhawy, F. H. & Trautmann, C. H. 1996. Estimation of in-situ test uncertainty. In Uncertainty in the Geologic Environment - From Theory to Practice (GSP 58), ASCE: 269 - 286, New York. 21

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

Lacasse, S., Nadim, F., Heg, K. & Gregersen, O. 2004. Risk assessment in geotechnical engineering: The importance of engineering judgement. In Proc. Advances in Geotechnical Engineering: The Skempton conference, (2): 856-867, London. Lind, N. C. 1971. Consistent partial safety factors. J. Struct. Div., ASCE, 97(ST6): 1651-1669. Low, B. K. & Phoon, K. K. 2002. Practical first-order reliability computations using spreadsheet. In Proc. Int. Conf. Probabilistics in GeoTechnics: Technical and Economic Risk Estimation, 30-46, Graz. Meyerhof, G. G. 1984. Safety factors and limit states analysis in geotechnical engineering. Can. Geotech. J., 21(1): 1-7. Ministry of Transportation and Communication 1979. Ontario Highway Bridge Design Code and Commentary (OHBDC1), 1st Ed., Downsview, Ontario. Ministry of Transportation Ontario 1983. Ontario Highway Bridge Design Code and Commentary (OHBDC2), 2nd Ed., Downsview, Ontario. Ministry of Transportation Ontario 1992. Ontario Highway Bridge Design Code and Commentary (OHBDC3), 3rd Ed., Downsview, Ontario. Moses, F. & Larrabee, R. D. 1988. Calibration of draft RP2A-LRFD for fixed platforms. In Proc. 20th Offshore Tech. Conf. (2): 171-180, Houston. National Research Council of Canada 1995. National Building Code of Canada (NBCC), 11th Ed., Ottawa. Nowak, A. J. & Lind, N. C. 1979. Practical bridge code calibration. In Proc. Specialty Conf. Prob. Mech. Struct. Reliability, 181-185, Tucson. Olson, R. E. and Dennis, N. D. 1982. Review and compilation of pile test results, axial pile capacity. Geotechnical Engineering Report CR83-4, Department of Civil Engineering, University of Texas, Austin. Ontario Hydro 1985. Limit state design of foundations - Design guide Part II, Deep foundations, Report no. 85122. Orr, T. L. L. 2002. Eurocode 7 a code for harmonized geotechnical design. In Proc. Int. Workshop on Foundation Design Codes and Soil Investigation in view of International Harmonization and Performance Based Design, 3 15, Tokyo. Ovesen, N. K. 1989. General report/discussion session 30: Codes and standards. In Proc. 12th Int. Conf. Soil Mech. Fdn. Engrg. (4): 2751-2764, Rio De Janeiro. Ovesen, N. K. 2002. Limit state design the Danish experience. In Proc. International Workshop on Foundation Design Codes and Soil Investigation in view of International Harmonization and Performance Based Design, 107 116, Tokyo. Paikowsky, S. G. & Stenersen, K. L. 2000. Performance of the dynamic methods, their controlling parameters and deep foundation specifications. In Proc. 6th Int. Conf. Appl. of Stress-Wave Theory to Piles: 281-304, So Paulo. Paikowsky, S. G. 2002. Load & Resistance Factor Design (LRFD) for deep foundations. NCHRP Report 24-17, Transportation Research Board, Washington. Phoon, K. K., Kulhawy, F. H. & Grigoriu, M. D. 1995. Reliability-based design of foundations for transmission line structures. Report TR-105000, Electric Power Research Institute, Palo Alto. (http://www.eng.nus.edu.sg/civil/aboutus/facultystaff/cvepkk/reliability.html) Phoon, K. K. & Kulhawy, F. H. 1999a. Characterization of geotechnical variability. Can. Geotech. J. 36(4): 612-624. Phoon, K. K. & Kulhawy, F. H. 1999b. Evaluation of geotechnical property variability. Can. Geotech. J. 36(4): 625-639. Phoon, K. K. & Kulhawy, F. H. 2002a. EPRI study on LRFD and MRFD for transmission line structure foundations. In Proc. Int. Workshop on Foundation Design Codes & Soil Investigation in view of International Harmonization & Performance Based Design, 253 261, Tokyo. Phoon, K. K. & Kulhawy, F. H. 2002b. Drilled shaft design for transmission line structure using LRFD and MRFD. In Deep Foundations 2002 (GSP116), ASCE, 1006 1017, New York.

22

Special lecture for Korean Geotechnical Society, Seoul, 9 July 2004 (KK Phoon)

Phoon, K. K. & Kulhawy, F. H. 2003. Evaluation of model uncertainties for reliability-based foundation design. In Proc. 9th Int. Conf. Applications of Statistics and Probability in Civil Engineering, (2): 13511356, San Francisco. Phoon, K. K., Becker, D. E., Kulhawy, F. H., Honjo, Y., Ovesen, N. K. & Lo, S. R. 2003a. Why consider reliability analysis in geotechnical limit state design? In. Proc. Int. Workshop on Limit State Design in Geotech. Engrg. Practice (LSD2003), Massachusetts Institute of Technology, Cambridge (CDROM). Phoon, K. K. & Kulhawy, F. H. & Grigoriu, M. D. 2003b. Development of a reliability-based design framework for transmission line structure foundations. J. Geotech. Geoenviron. Engrg., ASCE, 129(9): 798-806. Phoon, K. K. & Kulhawy, F. H. & Grigoriu, M. D. 2003c. Multiple resistance factor design (MRFD) for spread foundations. J. Geotech. Geoenviron. Engrg., ASCE, 129(9): 807-818. Phoon, K. K., Liu, S. L. & Chow, Y. K. 2003d. Estimation of model uncertainties for reliability-based design of cantilever walls in sand. In. Proc. Int. Workshop on Limit State Design in Geotech. Engrg. Practice (LSD2003), Massachusetts Institute of Technology, Cambridge (CDROM). Phoon, K. K. & Kulhawy, F. H. 2004. Reliability-based design of drilled shafts under undrained lateralmoment loading. In Geo-Trans 2004, ASCE, New York (in press). Pugsley, A. 1955. Report on structural safety. Structural Engineer, 33(5): 141-149. Rackwitz, R. & Fiessler, B. 1978. Structural reliability under combined random load sequences. Comp. & Struct., 9: 484-494. Ravindra, M. K. & Galambos, T. V. 1978. Load and resistance factor design for steel. J. Struct. Div., ASCE, 104(ST9): 1427-1441. Reese, L. C. 1958. Discussion of Soil modulus for laterally loaded piles. Transactions, ASCE, 123: 1071-1074. Reese, L. C., Cox, W. R. & Coop, F. D. 1974. Analysis of laterally loaded piles in sand. In Proc. 6th Offshore Tech. Conf. (2): 473 - 483, Houston. Rojiani, K. B., Ooi, P. S. K. & Tan, C. K. 1991. Calibration of load factor design code for highway bridge foundations. In Proc. Geotech. Engrg. Congress (GSP27), ASCE, 1353-1364, New York. Semple, R. M. 1981. Partial coefficients design in geotechnics. Ground Engrg., 14(6): 47 - 48. Sidi, I. D. 1986. Probabilistic prediction of friction pile capacities. PhD Thesis, University of Illinois, Urbana-Champaign. Simpson, B., Pappin, J. W., & Croft, D. D. 1981. An approach to limit state calculations in geotechnics. Ground Engrg., 14(6): 21 - 28. Task Committee on Structural Loadings 1991. Guidelines for electrical transmission line structural loading. Manual and Report on Engineering Practice 74, ASCE, New York. Valsangkar, A. J. & Schriver, A. B. 1991. Partial and total factors of safety in anchored sheet pile design. Can. Geotech. J., 28(6): 812-817. Wen, Y. K. 2000. Reliability and performance based design. In Proc. 8th ASCE Specialty Conf on Prob. Mech. Struct. Reliability: PMC2000-YKW on CDROM, University of Notre Dame. Whitman, R. V. 1984. Evaluating calculated risk in geotechnical engineering. J. Geotech. Engrg., ASCE, 110(2): 145-188. Withiam, J. L. 2003. Implementation of the AASHTO LRFD Bridge Design Specifications for Substructure Design. In. Proc. Int. Workshop on Limit State Design in Geotech. Engrg. Practice (LSD2003), Massachusetts Institute of Technology, Cambridge (CDROM). Yoon, G. L. & O'Neill, M. W. 1997. Resistance factors for single driven piles from experiments. Research Record 1569, Transportation Research Board, Washington, 47-54. U.S. Army Corps of Engineers 1997. Engineering and design, introduction to probability and reliability methods for use in geotechnical engineering, Engr. Tech. Letter No. 1110-2-547, Department of the Army, Washington.

23