Ensemble meta-classifiers and artificial neural networks for liquefaction potential and severity index prediction

Liquefaction-induced ground damage remains a critical challenge in seismically active regions, particularly following large-magnitude earthquakes. While conventional liquefaction assessments are typically limited to susceptibility evaluation or single-index analysis, integrated and scenario-based prediction of depthdependent liquefaction hazards remains insufficiently explored. This study presents a unified machine learning–based framework for the simultaneous prediction of the Liquefaction Potential Index (LPI) and Liquefaction Severity Index (LSI), addressing an important gap in existing liquefaction modeling studies. A comprehensive dataset comprising 320 borehole locations (160 liquefaction-observed and 160 non-observed sites) was analyzed under five earthquake scenarios with moment magnitudes (Mw) ranging from 6.0 to 8.0 and corresponding peak ground acceleration (PGA) values between 0.2 g and 0.6 g. Geotechnical and seismic input parameters—including groundwater level (GWL), average shear wave velocity (Vs30), Standard Penetration Test (SPT) values, cyclic resistance ratio (CRR), cyclic stress ratio (CSR), effective stress (o), Mw, and PGA—were used to compute LPI and LSI based on established methodologies. Predictive modeling was performed using an Artificial Neural Network (ANN) based on the Multilayer Perceptron (MLP) architecture and four ensemble learning techniques: Additive Regression, Bagging, Stacking, and Voting. Model performance was evaluated using 10-fold cross-validation and multiple statistical metrics, including the coefficient of determination (R2), RMSE, MAE, RAE, P/R ratio, and ANOVA. Among all models, Additive Regression exhibited superior predictive performance, achieving R2 values of 0.97 for LPI and 0.98 for LSI, and consistently outperforming the other approaches. Taylor diagram and error distribution analyses further confirmed the robustness and reliability of the proposed framework. The results demonstrate that the proposed approach provides a computationally efficient and engineering-relevant tool for scenario-based regional liquefaction hazard assessment, supporting rapid post-earthquake screening and informed decision-making in earthquake-prone areas