Smart ensemble modeling for multiclass lateral‑spreading prediction

Liquefaction-induced lateral spreading is a major geohazard that poses serious risks to infrastructure in seismically active regions, and although machine learning techniques have been increasingly applied to this problem, existing studies have predominantly formulated it using regression or binary classification frameworks. This paper introduces a novel explainable Hybrid Ensemble Network for Lateral Spreading (HEN-LS) for the multiclass prediction of liquefaction-induced lateral spreading. Leveraging a dataset of more than 6,000 seismic sites labeled across four displacement classes (very low, low, medium, and high), Optuna-tuned LightGBM, XGBoost, and TabNet base learners were developed. The probabilistic outputs of these learners were fused using a deep neural network meta-learner to construct the HEN-LS model. Model evaluation using confusion matrices and performance metrics demonstrated an overall accuracy of 92.1%, precision of 85.8%, recall of 89.3%, F1-score of 87.3%, Matthews correlation coefficient of 0.864, and a Brier score of 0.158. Permutation testing, Cochran’s Q test, and Bowker’s symmetry test (all p < 0.001) confirmed statistically significant improvements over individual base models. Explainability analyses using SHAP values and TabNet attention masks identified peak ground acceleration, river proximity, elevation, and groundwater depth as the dominant predictors, with class-specific attention patterns reflecting their varying influence across displacement categories. Confusion-matrix inspection further showed that misclassification rates for each class were consistently lower than those of individual models, highlighting the balanced performance of HEN-LS across all risk levels. Overall, HEN-LS combines high predictive accuracy with transparent decision-making at modest computational cost, offering a scalable tool for geotechnical hazard evaluation.