Rapid prediction of roof weighting behavior in karst underground coal mines using simulation-based deep learning

Karst peak-cluster mining poses a challenging roof-control problem because strong topographic relief induces highly heterogeneous overburden stress redistribution. At Fa'er Coal Mine, unexpected weighting events overloaded hydraulic supports by more than 20% and caused 72-h production interruptions. Conventional approaches remain too slow for rapid multi-scenario design screening. We therefore develop a physics-informed Mamba–CNN framework that predicts pressure class and weighting interval directly from FLAC3D contour images and physical-similarity observations. The model achieves 89.3% classification accuracy, 1.58 m RMSE, and a field MAE of 0.42 m (2.5%) against eight documented weighting events. Physics-informed constraints enforce consistency with established rock-mechanics relationships and eliminate physically inconsistent predictions. Inference requires 22 ms per scenario, enabling rapid post-simulation parameter extraction and screening across pre-generated simulation libraries. Field validation further shows that the predicted interval supports correct selection of the ZY4400 hydraulic support system, demonstrating practical value for proactive roof-control design in topographically complex mines.