Application of Statistical Machine Learning for Reliability Assessment of Liquefaction Analysis with Unobservable Multivariate Parameters and Identification of Challenges

Abstract

This study examines the statistical estimation methods for multivariate analysis parameters in Effective Stress Dynamic Analysis (LIQCA). The constitutive model used in LIQCA include several fitting parameters that cannot be directly observed in the field. he accuracy of setting these fitting parameters often depends on the analyst’s experience. Therefore, to objectively describe the reliability of the analysis, it is necessary to clarify the process of setting these parameters. In this study, we collected and organized the parameters and soil test results used in the parameter-setting process from 27 LIQCA analysis cases conducted by experts in this field, and constructed a database. Using the statistical information of parameter sets obtained from the database as prior information, simple Bayesian linear regression was applied as a reference solution. Furthermore, clustering was performed on the database, and linear regression using hierarchical Bayesian methods and sampling using similarity evaluation were applied to discuss the impact of the estimation results on the uncertainty of the numerical analysis results. Based on these considerations, this paper proposes an effective framework for utilizing the experience of analysts in setting material parameters and discusses the problems in promoting these database approaches.