Comparative Evaluation of Machine Learning Models for Discontinuity-Controlled Block Stability in Underground Caverns

Discontinuity-controlled rock masses in underground caverns are prone to block formation and instabilities during construction, motivating rapid tools for stability feedback analysis. Fast screening of discontinuity-controlled block hazards in underground caverns is addressed through a physics-consistent machine learning framework for identifying block formation and assessing instability conditional on formation. Cavern geometry, rock mass properties, and multiple joint sets are parametrically encoded, and physically and geometrically consistent data augmentation is performed. Three-dimensional discrete element batch simulations provide automated labels for block formation (yB) and instability (yU) (FoS < 1), forming a training dataset. Twenty-three raw features with a 66-dimensional engineered feature set are compared and multiple classifiers using PR/ROC curves, confusion matrices, and a BAcc-based threshold strategy are evaluated. Compared with raw inputs, the engineered features generally improve AUC-based ranking and balanced discrimination, especially for the RF model, although threshold-dependent recall and F1 trade-offs are observed for some learners. Random forests show consistently robust results. Benchmarking against an independent engineering reference based on stereographic projection and limit-equilibrium analysis gives 91% agreement for block-formation identification and 85.2% agreement for conditional instability identification. The trained model is integrated into an engineering platform to support batch screening of three- and four-plane combinations with 3D visualization outputs.