Correlation analysis between the collapsibility of loess and physicomechanical indices in Longdong region and prediction of the collapsibility coefficient based on machine learning.

Against the backdrop of booming engineering construction in loess areas, achieving rapid and accurate prediction of the loess collapsibility coefficient is of great significance. This study takes the loess in the Longdong area as the research object. Based on analyzing the correlation between the collapsibility coefficient and various soil property indices, a rapid determination standard for the collapsibility of loess in this area is proposed. Prediction models between the loess collapsibility coefficient and void ratio, depth, saturation, and compression coefficient are established using three machine learning algorithms: random forest (RF), extreme gradient boosting (XGBoost), and support vector machine (SVM). The prediction results are subjected to interpretability analysis using the Shapley additive explanations (SHAP) technique. The research results show that loess with a void ratio less than 0.960 and a soil layer depth greater than 16 m can be used as the rapid determination standard for non-collapsible loess in this area. The XGBoost model has higher prediction accuracy and applicability, with coefficients of determination (R