Performance evaluation of machine learning algorithms for predicting liquefaction-induced lateral displacement

Accurate prediction of liquefaction-induced lateral displacement is essential for seismic risk assessment, resilient infrastructure design, and cost-effective mitigation. Such predictions are complex and cannot be reliably addressed using conventional analytical approaches. This research utilizes XGBoost, CatBoost and AdaBoost algorithms with 247 post-liquefaction in-situ free-face ground condition case studies to model and investigate liquefaction induced lateral displacements. The models are assessed according to the coefficient of determination (R²), coefficient of correlation (r), mean absolute error (MAE), mean squared error (MSE), root means square error (RMSE), and Root Mean Squared Error-Observations standard deviation Ratio (RSR) and Nash Sutcliffe Efficiency (NSE) coefficient. The develop models are compared with each other, and also to the Gaussian process regression, artificial neural network, Evolutionary polynomial regression and Multiple linear regression models described in the literature. The XGBoost model had the best prediction performance, with R² = 0.9905, r = 0.9952, MAE = 0.1491, MSE = 0.0485, RMSE = 0.2203, RSR = 0.0981, and NSE = 0.9904 for training and R² = 0.9251, r = 0.9618, MAE = 0.3642, MSE = 0.3723, RMSE = 0.6101, RSR = 0.278, and NSE = 0.9227 for testing results respectively. The rank score analysis confirmed XGBoost to be superior, as it reached the highest total score of 40. Sensitivity analysis further revealed that T₁₅ was the most sensitive parameter to output variable.