Predictive Modeling of Soil Liquefaction Potential Using Hybrid Deep Learning Algorithms

Abstract

Soil liquefaction remains one of the most destructive earthquake hazards, and its reliable prediction underpins seismic microzonation, foundation design and lifeline risk management. Semi-empirical stress-based procedures return threshold-type answers that are sensitive to input error, whereas stand-alone machine-learning models often lack calibrated probabilities and are rarely tested under distribution shift. This study develops and benchmarks a hybrid deep-learning framework, PSO-DNN-XGB, that couples a particle swarm optimisation (PSO)-tuned deep neural network (DNN) with a regularised gradient-boosting head: the last hidden layer of the DNN is concatenated with physically derived features and passed to XGBoost to yield calibrated liquefaction probabilities. The framework is evaluated on a semi-synthetic database of 1,500 SPT-based case histories generated from Boulanger–Idriss relationships with model and measurement uncertainty, and on an independent 350-case set with deliberate covariate shift. Against seven benchmarks, the hybrid attains a test accuracy of 0.876, F1 of 0.846, AUC of 0.948, MCC of 0.742 and Brier score of 0.092, and the highest external accuracy (0.897). It markedly improves on an untuned DNN (MCC 0.620) and is statistically on par with the deterministic reference, which is near-optimal by construction here. Ablation shows that PSO tuning and feature concatenation each contribute materially, and SHAP analysis identifies CSR7.5 and (N1)60cs as dominant drivers, consistent with engineering theory. The results are a controlled methodological benchmark that must be confirmed on field case-history databases.