Advancements in predicting soil liquefaction susceptibility: a comprehensive analysis of ensemble and deep learning approaches.

Liquefaction is a phenomenon that occurs when there is a loss of strength in wet and cohesionless soil due to higher pore water pressures, and as a result, the effective stress is reduced due to dynamic loading. A construction site should first investigate the site for liquefaction, and for analyzing liquefaction, the most accurate method should be selected, which provides the most accurate results. In this work, a detailed investigation is performed on the effectiveness of ensemble learning and deep learning (DL) models in assessing the liquefaction susceptibility of soil deposits from a large database consisting of cone penetration test (CPT) measurements and field liquefaction performance observations of historical earthquakes. The performance of the developed models is assessed via several comprehensive performance fitness error matrices (PFEMs), including precision, accuracy, recall, specificity, F1 score, MCC, BA, receiver operating characteristic (ROC) curve analysis, and area under the curve (AUC). Accuracy and validation loss curves were also plotted for all the proposed models. PFEMs are calculated, and a comparative study is performed for all the proposed methods. The BI-LSTM model has the highest accuracy, with 0.9791 in training and 0.8889 in testing, indicating strong predictive ability and good generalizability. LSTM follows closely with training and testing accuracies of 0.9433 and 0.8750, respectively, offering consistent performance. XGBoost also performs well, achieving 0.9194 in training and 0.8750 in testing, reflecting its robustness in handling complex patterns. In contrast, RF displays a significant discrepancy between the training (0.9373) and testing (0.8681) accuracies. Overall, BI-LSTM emerges as the most reliable model for assessing liquefaction potential, with LSTM, XGBoost and the RF also proving effective. Each model can offer unique strengths, with BI-LSTM and LSTM excelling at learning sequential dependencies, whereas XGBoost and RF provide powerful and often interpretable results from structured tabular data. This study advances the development of robust tools for assessing liquefaction hazards, thereby enhancing strategies for seismic risk mitigation.