Elsevier

Engineering Geology

Volume 203, 25 March 2016, Pages 191-203
Engineering Geology

Inter-region variability of Robertson and Wride method for liquefaction hazard analysis

https://doi.org/10.1016/j.enggeo.2015.12.024Get rights and content

Highlights

  • Method for characterizing the inter-region variability is presented.

  • The inter-region variability of Robertson and Wride Method is characterized.

  • The accuracy of the Robertson and Wride model is region dependent.

  • Region-specific FscPL relationships are derived for different regions.

  • Collecting more region-specific data helps reducing the inter-region variability.

Abstract

Semi-empirical models used for liquefaction potential evaluation are often developed based on lumping data from different regions around the world. The accuracy of a semi-empirical model developed in such a way may vary from one region to another, which has been seldom studied previously. Thus, in this paper, a probabilistic method is suggested to characterize the inter-region variability of the model bias factor of a semi-empirical model for liquefaction potential evaluation. As an illustration, the proposed approach is used to characterize the inter-region variability of the model bias factor of the Robertson and Wride model (called RW model for short herein). It is found that substantial inter-region variability exists in the model bias factor of the RW model. For a region not in the calibration database, the uncertainty associated with the model bias factor is significantly larger than that of a region in the calibration database. For regions in the calibration database, the uncertainty associated with the model bias factor in a region also changes with the amount and the type of the calibration data available in that region. It is thus important to collect region-specific performance data for model calibration to achieve accurate liquefaction potential evaluation. For ease of practical application, equations are provided for estimating the liquefaction probability based on the factor of safety (FOS) computed with the RW model for different regions. The suggested method can also be used to characterize the inter-region variability of other semi-empirical models and to derive region-specific relationships between FOS and liquefaction probability for these models.

Introduction

Due to the difficulties to obtain high quality soil samples, semi-empirical methods based on soil indexes measured with in-situ tests such as cone penetration test (CPT) are widely used to assess the liquefaction potential of soils. Such methods are generally developed based on a database that includes case histories derived from earthquakes in different regions around the world (e.g., Robertson and Wride, 1998, Cetin et al., 2002, Moss et al., 2006, Muduli et al., 2014). The semi-empirical methods are often applied universally around the world. The assumption behind such a practice is that the model uncertainty associated with an empirical model is homogenous for all regions. Such an assumption, however, may not be correct for certain cases. For instance, consider the Robertson and Wride model (Robertson and Wride, 1998, Robertson, 2009), one of the most widely used CPT-based models for evaluating the potential of soil liquefaction. For convenience of presentation, the Robertson and Wride method is referred to hereinafter as the RW model. Table 1 shows the post-earthquake investigation data from 11 major earthquakes, which are adapted from the database compiled in Ku et al. (2012). The factor of safety (FOS) computed based on the RW model is shown in the last column of Table 1. In principle, a soil should liquefy if the FOS against liquefaction is less than unity, and should not liquefy if the FOS is larger than unity. Based on this criterion, there are three out of 11 cases from the 1976 Tangshan earthquake zone in which the liquefaction phenomenon is wrongly predicted by the RW model. For the three wrongly predicted cases, the values of computed FOS are all smaller than unity but liquefaction did not occur, indicating that the RW model may tend to underestimate the actual FOS of soils from the Tangshan earthquake zone. For comparison, among the five cases from the 1980 Mexicali earthquake zone, four cases are wrongly predicted. For these wrongly predicted cases, the values of computed FOS are all larger than unity but soil liquefaction was observed, indicating the RW model may tend to overestimate the actual FOS of soils from the 1980 Mexicali earthquake zone. It seems that the characteristics of model uncertainty associated with the RW model could vary from one region to another, indicating the existence of inter-region variability about the applicability of the RW model.

The existence of inter-region variability in empirical relationships has been previously observed in other geotechnical engineering problems (e.g., Zhang et al., 2004). In foundation engineering, it is often quite typical to develop local empirical relationships to predict the pile capacity based on data in one region to avoid the adverse effect of inter-region variability (e.g., McVay et al., 2003, Reddy and Stuedlein, 2013). For liquefaction problems, however, such an idea is not feasible because post-earthquake soil liquefaction data are scarce and the data in one region are often insufficient to develop a local model. As such, lumping data from different regions for the development of empirical liquefaction potential evaluation models is necessary and the existence of inter-region variability may be unavoidable.

Research studies have been carried out to assess the model uncertainty associated with liquefaction potential evaluation models. Cetin et al. (2002) developed a model to evaluate the liquefaction potential of soils based on the standard penetration test (SPT) data considering the variability in soil parameters, and the model error of the suggested model is characterized during the model development process. Using a similar procedure, Moss et al. (2006) developed a liquefaction potential prediction model based on the CPT data with explicit consideration of model uncertainty. Huang et al. (2012) suggested a Bayesian network for characterizing the model uncertainty associated with a liquefaction model considering the uncertainties in the input soil parameters. A maximum likelihood-based method is employed in Ku et al. (2012) to characterize the model uncertainty associated with the RW model, and a relationship between the computed FOS and liquefaction probability was derived. In these studies it is assumed that the distributions of the model bias factors of different regions are identical, i.e., the model uncertainty is homogenous across regions. The issues of how to model and characterize the inter-region variability and how such variability affects the liquefaction potential evaluation have been seldom addressed in the literature.

The objective of this paper is to extend the work in Ku et al. (2012) and to suggest a method to characterize the inter-region variability of the accuracy of the RW model, through which the effect of inter-region variability of the model bias factor on liquefaction potential evaluation will also be studied. This paper is organized as follows. First, the probabilistic model for modeling the inter-region variability is suggested, followed by a Bayesian method for calibrating the suggested model. Then, the procedure is presented for comparing models based on different assumptions regarding the inter-region variability of the model bias factor. Finally, the inter-region variability of the model bias factor of the RW model is investigated with the proposed approach, and its impact on liquefaction potential evaluation is discussed. This paper provides useful insights on how to address the inter-region variability in the semi-empirical models and how to use such models appropriately for liquefaction potential evaluation. The method suggested in this paper can also be used to characterize the inter-region variability of other semi-empirical methods for liquefaction potential evaluation.

Section snippets

Probabilistic model

Let Fsc denote the FOS calculated by the RW model, and let Fsa denote the actual FOS. Due to the existence of model uncertainty, Fsc may not be exactly Fsa. Suppose Fsc can be related to Fsa via a model bias factor c as follows (e.g., Ang and Tang, 1984)Fsa=cFsc

The model bias factor c can be assumed to follow the lognormal distribution (e.g., Juang et al., 2004, Huang et al., 2012). Let μci and σci denote the mean and the standard deviation of the model bias factor at the ith region. As

Ranking of competing models

As will be seen later in this study, based on the same calibration data, it is possible to construct several probabilistic models depending on whether the inter-region variability is modeled and how it is modeled. Yet, the predictions from these models could be quite different. In such a case, how to compare the validity of different models is important. In this study, different models are compared using Bayes' theorem, where model probability is used to measure the plausibility of different

Calibration database

Ku et al. (2012) compiled a database comprising 165 cases, which are derived from 16 earthquakes around the world based on previous studies of Moss et al. (2006), (2011); Robertson (2009). For the purpose of calibrating the inter-region variability, this database is reviewed and earthquakes with less than 4 cases are excluded for the argument that when the number of cases is small, the observed data provide little information about the model bias factor in that region. The new database in this

Summary and conclusions

The research reported in this paper and findings from this paper can be summarized as follows:

  • (1)

    A probabilistic model is suggested to model the inter-region variability associated with the model bias factor of the Robertson and Wride (RW) model for liquefaction potential evaluation, in which the mean values of the model bias factor at different regions are assumed different but follow a common distribution. A method based on multivariate Bayesian updating has been suggested for model calibration

Acknowledgment

The first author is grateful to the support from the National 973 Basic Research Program of China (2014CB049100), the Shanghai Rising-star Program (15QA1403800), and the Fundamental Research Funds for Central Universities.

References (31)

  • A. Gelman et al.

    Bayesian Data Analysis

    (2013)
  • R.B. Gilbert

    First-order, Second-moment Bayesian Method for Data Analysis in Decision Making

    (1999)
  • H.W. Huang et al.

    Bayesian network for characterizing model uncertainty of liquefaction potential evaluation models

    KSCE J. Civ. Eng.

    (2012)
  • S. Jackman

    Bayesian Analysis for the Social Sciences

    (2009)
  • C.H. Juang et al.

    Maximum likelihood principle and its application in soil liquefaction assessment

  • Cited by (27)

    • Bayesian partial pooling to reduce uncertainty in overcoring rock stress estimation

      2023, Journal of Rock Mechanics and Geotechnical Engineering
    • Model calibration: A hierarchical Bayesian approach

      2023, Probabilistic Engineering Mechanics
      Citation Excerpt :

      Interestingly, hierarchical models are absent as a general approach to data analysis from the decades-long history of application of Bayesian techniques in geotechnical engineering (e.g. see review papers by Zhang et al. [12] and Baecher [13]). Zhang et al. [14, 15] are among early examples of using hierarchical models for systematic modelling of between-site variations. Bozorgzadeh et al. [16] and Bozorgzadeh and Bathurst [8] provide relatively accessible introductions to hierarchical modelling as a general means of approaching geotechnical engineering data.

    • Hierarchical Bayesian modelling of quasi-region-specific soil porosity

      2022, Ocean Engineering
      Citation Excerpt :

      It should be noted that the parameters of the I–n regressions (ai and bi) at the bottom level are conditionally modelled based on μa, σa, μb and σb in the upper level, where μa and μb are the mean values of the slopes and intercepts, respectively. In addition, the hyperparameters σa and σb represent the variabilities between different regions, which is named inter-regional variability (Zhang et al., 2014, 2016), while these four parameters μa, σa, μb and σb are called hyperparameters (Gelman and Hill, 2006). In other words, a model with hyperparameters gives rise to sub-models with parameters ai and bi, which then produce the observed data (Gelman et al., 2013).

    View all citing articles on Scopus
    View full text