Data-Driven Explainable Machine learning for Rockburst Hazard and Seismic Resilience with 3D Fragility Surfaces

Rockbursts are critical geohazards in deep underground mining and tunneling and can cause catastrophic structural failures, severe safety risks, and substantial economic losses. Owing to their nonlinear and spatiotemporally variable behavior, conventional microseismic monitoring approaches often provide limited predictive capability. This study proposes a data-driven machine learning (ML) framework to model rockburst fragility surfaces and assess seismic resilience under complex geological conditions. The workflow begins with preprocessing of in-situ monitoring data, including dimensionality reduction, feature engineering, and normalization to enhance data quality and relevance. Multiple ML algorithms, including XGBoost, Random Forest, LightGBM, and CatBoost, are trained using labeled datasets representing moderate-to-severe rockburst events. To improve robustness and generalization, an ensemble voting classifier is developed, with hyperparameters optimized using grid search and evolutionary optimization. The framework maintains strong predictive consistency even for imbalanced datasets. Feature importance analysis identifies seismic energy release and ground stress as the dominant predictors of rockburst occurrence. Using these insights, the model generates three-dimensional fragility surfaces that characterize structural vulnerability under varying seismic loads, along with probabilistic fragility curves that delineate risk zones to support early-warning systems and real-time operational decision-making. By integrating physics-informed assessment with explainable ML, the proposed approach improves interpretability while meeting practical engineering requirements. Overall, this study provides a scalable, high-performance framework for seismic resilience analysis, enabling improved hazard prediction, risk management, and the design of safer underground systems in complex geomechanical environments.