Micronano Pore Structure Characterization and Machine Learning for Methane Adsorption Prediction in Mining Coal

Abstract

Coal pore structures evolve continuously during mining-induced deformation, but their nonlinear effects on methane adsorption remain difficult to quantify using conventional empirical models. In this study, pore structure evolution in coal samples at different mining stages was characterized using scanning electron microscopy (SEM) and low-temperature N2 adsorption. Methane adsorption isotherms were measured at each mining stage using a high-pressure volumetric method. To capture the nonlinear relationship between pore structure parameters and methane adsorption capacity, extreme gradient boosting (XGBoost), random forest (RF), and support vector machine regression (SVM) models were optimized using the Kepler and RIME algorithms. Among these models, RIME-SVM achieved the best prediction performance, with R2 greater than 0.9200 and mean absolute percentage error (MAPE) lower than 0.1890. Sensitivity analysis showed that mining Stages III–V had the strongest influence on Stage VI methane adsorption, whereas Stage I had little effect. In addition, a cubic polynomial model described the interstage variation in adsorption capacity, most R2 values greater than 0.8200, further linking coal mining disturbance with methane adsorption evolution. These results provide a quantitative method for predicting methane adsorption changes in mining coal and can support coalbed methane extraction and mine gas control.