ZHANG Ling, CHEN Yun-hao, WANG Wen-tao, LUO Biao, LIU Chang-jie, TAN Jing-peng
The long-term stability of mudstone coarse-grained soil subgrades is closely related to their strength. Dynamic-static triaxial tests were conducted on mudstone coarse-grained soils with varying degrees of erosion to accurately predict their strength under the effects of penetrating erosion and cyclic loading. A peak strength database containing 300 datasets was established. Based on this database, six machine-learning methods-random forest (RF), gated recurrent unit (GRU), k-nearest neighbors (KNN), backpropagation neural network (BPNN), extreme gradient boosting (XGBoost), and Gaussian process regression (GPR)-were employed to predict the peak strength. The models were comprehensively evaluated using metrics, such as determination coefficient R2, root mean square error (RMSE), mean absolute error (MAE), 95% uncertainty interval (U95), and weighted mean differences (WMD). The results indicate that six factors-penetrating intensity (p0), loading frequency (f), dynamic confining pressure (σ3c), dynamic deviatoric stress (σd), loading cycle count (N), and static confining pressure (σ3s)-are the primary contributors to the peak strength (σ1-σ3)max. Among the methods tested, the XGBoost model demonstrates the highest prediction accuracy and lowest uncertainty and is recommended as the optimal method. Sensitivity analysis reveals that penetrating intensity (p0) makes the greatest contribution to (σ1-σ3)max (with a weight of 30.28%), followed by loading cycle count (N) (27.81%) and static confining pressure (σ3s) (21.91%). Monotonicity analysis shows that with increasing p0, σd, f, and N, (σ1-σ3)max generally decreases. Notably, when p0 ranges from 0 to 40 kPa, the decrease in (σ1-σ3)max is less significant. Thus, coarse particles dominate the mechanical behavior, and the influence of erosion on strength is minimal during this range.
