Accurate and generalizable soil liquefaction prediction model based on the CatBoost algorithm

Abstract

Accurate prediction of soil liquefaction is important for preventing geological disasters. Soil liquefaction prediction models based on machine learning algorithms are efficient and accurate; however, some models fail to achieve highly precise soil liquefaction predictions in certain areas because of poor generalizability, which limits their applicability. Thus, a soil liquefaction prediction model was constructed using the CatBoost (CB) algorithm to support categorical features. The model was trained using standard liquefaction datasets from domestic and foreign sources and was optimized with Optuna hyperparameters. Additionally, the model was evaluated using five evaluation metrics and its performance was compared to that of other models that use multi-layer perceptron, support vector machine, random forest, and XGBoost algorithms. Finally, the prediction capability of the model was verified using three case studies. Experimental results demonstrated that the CB-based model generated more accurate soil liquefaction predictions than other comparison models and maintained their performance. Hence, the proposed model accurately predicts soil liquefaction and offers strong generalizability, demonstrating the potential to contribute toward the prevention and control of soil liquefaction in engineering projects, and toward ensuring the safety and stability of structures built on or near liquefiable soils.