Predicting Soil Liquefaction Potential Using XGBoost Algorithm with Bayesian Hyperparameters’ Optimization

Liquefaction potential is one of the most important measures when assessing soil liquefaction hazard. With advancements in prediction tools, over the past decade, many researchers have started applying machine learning algorithms to obtain a more accurate assessment of soil liquefaction potential. The adopted algorithms often require rigorous computational efforts to select the optimum set of parameters (referred to as hyperparameters) that leads to a reliable and accurate prediction. This paper presents a hyperparameter tuning method based on a probabilistic technique known as Bayesian learning, in which the optimum set of hyperparameters is selected providing a high prediction accuracy with less computational effort. For illustration purposes, a dataset of 170 case histories of CPT-based soil liquefaction from previous earthquakes is adopted. The XGBoost algorithm, which is known for providing high accuracy models for regression and classification purposes, is adopted herein for prediction of liquefaction potential. Bayesian hyperparameter optimization method is adopted for selecting the optimum set of parameters. The liquefaction potential prediction obtained through this approach is compared with other machine learning models developed by other researchers on the same dataset. The Bayesian hyperparameter optimization is shown to improve the prediction accuracy.