Application of different machine learning algorithms in post-earthquake failure probability assessment of underground structures

Abstract

Seismic vulnerability analysis of underground structures is critical for ensuring the resilience of urban infrastructure against earthquake hazards. However, traditional finite Element method-based cloud models, face challenges due to uncertainties from both aleatory and epistemic aspects, leading to exponentially increasing computational demands. To address these limitations, this study proposes a novel framework integrating the cloud model with machine learning (ML) algorithms to enhance the efficiency of seismic fragility analysis. Specifically, three supervised ML algorithms consisting of Back-Propagation Neural Network (BPNN), Support Vector Regression (SVR), and Random Forest Regression (RF), are employed to predict structural seismic responses, using FEM-derived results as the training dataset. Through rigorous feature selection, data preprocessing, dataset partitioning, and hyperparameter optimization via grid search and cross-validation, robust ML models are developed. Consequently, seismic vulnerability curves are constructed using logarithmic linear regression to correlate ground motion intensity measures (IMs) with structural damage measures (DMs). By comparing these ML-derived curves with finite element analysis results, the Wasserstein distance reveals discrepancies below 0.045, with RF demonstrating the smallest deviations (below 0.025) and superior stability across four full damage states of the subway stations. These findings confirm the reliability and computational efficiency of the proposed ML-based approach, offering a practical alternative for seismic fragility assessment of subway stations and advancing seismic-resistant design.