Interpretable Random Forest Framework for Predicting Triaxial Shear Strength Parameters of Gypseous Soils Using Modified Collapse Test Indicators

Abstract

Gypseous soils exhibit moisture-sensitive behavior because gypsum may act as a cementing agent under natural conditions but dissolves upon wetting, causing collapse and shear strength degradation. This study develops an interpretable Random Forest Regression framework to predict the shear strength parameters of collapsible gypseous soils by integrating UU/CU triaxial test results with Modified Collapse Test (MCT) data. Artificial gypseous soils were prepared with gypsum contents of 20%, 40%, and 60% and compacted at dry unit weights of 14, 15, and 16  The input variables were gypsum content, dry unit weight, initial water content, degree of saturation, initial void ratio, CP-MCT, and test condition, while the outputs were cohesion (c) and internal friction angle (φ). Triaxial stress-strain variables were excluded to avoid data leakage. A constrained Monte Carlo-based augmentation procedure was used only to improve training stability within the experimental domain, while model performance was interpreted with reference to the original experimental observations. The RF model incorporating CP-MCT achieved R² = 0.777 for c and R² = 0.869 for φ. Feature importance analysis identified testing condition as the dominant factor, confirming the strong influence of saturation on shear strength degradation. Although the inclusion of CP-MCT did not improve numerical accuracy relative to the baseline model without CP-MCT, it enhanced the physical interpretability of the framework by explicitly linking collapse susceptibility with strength loss. Because the experimental database is limited in size, the proposed framework should be regarded as a preliminary decision-support and interpretation tool rather than a replacement for laboratory testing or final design verification.