Scholarlink

<p>An accurate assessment of soil liquefaction probability is crucial for risk-informed geotechnical design; however, the traditional Monte Carlo Simulation (MCS) methods are computationally inefficient for rare event probabilities. This study presents the subset simulation (SS) approach as a more efficient alternative for estimating low-probability liquefaction events within a simplified framework that integrates a state parameter (<i>ψ</i>)-based cyclic resistance ratio (CRR) model. The integration of the <i>ψ</i>, derived from critical state soil mechanics, enhances the model’s ability to account for the effects of relative density and effective stress. In addition, we develop and evaluate three machine learning (ML) models to predict SS-based <i>P</i><sub><i>L</i></sub>. This study uses the Optuna framework for hyperparameter optimisation, demonstrating its efficiency in DNN models where exhaustive methods, such as grid search, are either impossible or computationally inefficient. The results demonstrate the robustness of the SS method in predicting small failure probabilities in liquefaction analysis. A logistic mapping function with a coefficient of determination (<i>R</i><sup>2</sup> = 0.97) was developed to relate factor of safety to probabilistic liquefaction estimates. Among the ML models evaluated, the DNN model exhibited enhanced predictive accuracy and generalisation capability relative to XGB and CatBoost, with an <i>R</i><sup><i>2</i></sup> of 0.9303 and an <i>RMSE</i> of 0.0794 on the test dataset. Finally, global sensitivity analysis using Sobol’ indices quantifies the influence of input variables on models’ predictions. This study contributes to the application of subset simulation in valuable risk assessment in geotechnical structure design and advocates for deep learning-based models in liquefaction evaluation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

State Parameter-Based Soil Liquefaction Assessment Using Subset Simulation and Machine Learning Approaches

  • 期刊论文

  • 发表日期：

    2025年7月24日

查看全文
  • Kishan Kumar,
  • Pijush Samui,
  • S. S. Choudhary,
  • Edy Tonnizam Mohamad

摘要

An accurate assessment of soil liquefaction probability is crucial for risk-informed geotechnical design; however, the traditional Monte Carlo Simulation (MCS) methods are computationally inefficient for rare event probabilities. This study presents the subset simulation (SS) approach as a more efficient alternative for estimating low-probability liquefaction events within a simplified framework that integrates a state parameter (ψ)-based cyclic resistance ratio (CRR) model. The integration of the ψ, derived from critical state soil mechanics, enhances the model’s ability to account for the effects of relative density and effective stress. In addition, we develop and evaluate three machine learning (ML) models to predict SS-based PL. This study uses the Optuna framework for hyperparameter optimisation, demonstrating its efficiency in DNN models where exhaustive methods, such as grid search, are either impossible or computationally inefficient. The results demonstrate the robustness of the SS method in predicting small failure probabilities in liquefaction analysis. A logistic mapping function with a coefficient of determination (R2 = 0.97) was developed to relate factor of safety to probabilistic liquefaction estimates. Among the ML models evaluated, the DNN model exhibited enhanced predictive accuracy and generalisation capability relative to XGB and CatBoost, with an R2 of 0.9303 and an RMSE of 0.0794 on the test dataset. Finally, global sensitivity analysis using Sobol’ indices quantifies the influence of input variables on models’ predictions. This study contributes to the application of subset simulation in valuable risk assessment in geotechnical structure design and advocates for deep learning-based models in liquefaction evaluation.