Metaheuristic-Driven Deep Learning Framework for Soil Liquefaction Assessment

Soil liquefaction in saturated sandy deposits remains a major concern in seismically active regions. This study presents a metaheuristic-driven deep learning framework for liquefaction assessment using a dataset of 512 boreholes collected across Mazandaran Province, Northern Iran. The dataset consists of 316 non-liquefied and 196 liquefied cases derived from field observations and verified geotechnical parameters. The proposed framework integrates Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) architectures, whose hyperparameters are optimized using the Grey Wolf Optimizer (GWO) and the Whale Optimization Algorithm (WOA). Model performance was evaluated using stratified repeated train–test experiments (10 iterations) to ensure reproducibility. The ensemble model achieved an average accuracy of 94.5%, a correlation coefficient of 0.93, and AUC values above 0.95, outperforming all individual models and conventional liquefaction evaluation methods. SHAP-based analysis identified SPT blow count and peak ground acceleration as the most influential variables. The proposed framework demonstrates strong robustness and practical applicability for regional seismic risk assessment and can be extended to other geotechnical prediction tasks.