A data-driven framework for predicting mechanical properties of coal–rock composites in deep mining
Abstract
The depletion of shallow coal resources necessitates the advancement of deep mining operations, where accurate prediction of coal–rock composite mechanical behavior is critical for disaster prevention. This study systematically develops and evaluates a machine learning (ML) framework for predicting key mechanical properties. A dataset of 162 experimental results was collected, incorporating ten input features such as densities, elastic modulus, uniaxial compressive strengths (UCSs), and key ratio parameters of coal–rock components. Five base models, namely backpropagation neural network (BPNN), extreme gradient boosting (XGBoost), support vector machine (SVM), extreme learning machine (ELM), and random forest (RF), were optimized using the sparrow search algorithm (SSA). The results show that the SSA–SVM model achieved the best prediction accuracy on test data, with determination coefficients of 0.90 for composite UCS (Urc) and 0.85 for elastic modulus (Erc). Feature importance analysis highlighted the significant influence of coal proportion (Rc) and strength ratios (Ru) on composite mechanical performance. As Rc increases, the overall composite strength decreases. Higher strength ratios also alter load-transfer efficiency between coal and rock and increase the risk of interfacial failure. This study provides data-driven insights for designing safer underground support systems in deep mines and mitigating mining disasters.