AI-enhanced geomechanical modeling: A dynamic mining land subsidence prediction framework with geology-triggered rebalancing

Mining land subsidence (MLS) poses a serious threat to the infrastructure and ecological security of mining areas. Traditional pure data-driven or empirical models often struggle to overcome the spatiotemporal heterogeneity of the subsidence process and lack the ability to adapt dynamic parameters in accordance with physical laws. Therefore, this study proposes a physics-informed hybrid prediction framework that achieves high-precision dynamic prediction of MLS through the deep integration of traditional mining rock mechanics mechanisms and modern artificial intelligence. Firstly, the swarm intelligence algorithm was collaboratively improved through Latin hypercube sampling, an improved Sigmoid function, sine and cosine oscillation adjustment, and a hybrid differential evolution strategy. The CEC2017 test demonstrated that the convergence accuracy and geological adaptability were significantly enhanced after the improvement. Second, a physics-constrained geology-triggered rebalancing (GTR) mechanism is proposed. Directly triggered by actual abrupt changes in physical subsidence, it executes dynamic weight transfers among 12 geomechanical parameters, including those of the probability integration method, accurately reproducing the evolution of the dominant mechanical mechanisms across the three stages of rock strata: elastic deformation, accelerated collapse, and residual adjustment. The results show that compared with traditional methods, the prediction error was reduced by 35%, and the dynamic parameter optimization reduced the crack rate of surface buildings in the Tangkou mining area by 66%. This study establishes a cutting-edge interdisciplinary paradigm that embeds traditional geomechanical theories into AI algorithms, providing scientific support for the prevention and control of geological hazards in deep underground mining.