Machine Learning-Based Prediction of Cumulative Dynamic Strain of Malan Loess under Wetting Conditions in Zhengzhou Area

To investigate the development of cumulative dynamic strain in Zhengzhou loess under wetting conditions and to develop a corresponding predictive framework, this study integrates dynamic triaxial testing with machine learning modeling. Experimental results indicate that, under cyclic dynamic loading, the cumulative dynamic strain of Zhengzhou loess exhibits a typical three-stage development pattern characterized by a rapid initial growth, a relatively stable intermediate accumulation, and an accelerated deformation in the late stage. Under identical confining pressure and dynamic stress amplitudes, the cumulative dynamic strain of saturated specimens is markedly larger than that of wet and natural-state specimens. An increase in confining pressure suppresses the development of cumulative dynamic strain, whereas higher dynamic stress amplitudes and a greater number of loading cycles substantially intensify deformation accumulation. Utilizing wetting-induced water content, confining pressure, dynamic stress amplitude, and loading cycles as inputs, eight machine learning algorithms were comparatively evaluated, and the Gradient Boosting Regression (GBR) model achieves superior predictive accuracy. Feature importance analysis based on SHAP values indicates that the relative contributions of the input variables to cumulative dynamic strain development are consistent with the established mechanical mechanisms of wetting-induced weakening, stress-driven deformation, and loading-history accumulation in loess.