Integration of computational algorithms in enhancing geological modelling and drilling efficiency in coal mine exploration

Coal calorific value calculation is crucial for power planning and mine exploration. Traditional lab techniques and empirical calculations often fail with high variability or limited data. We developed a machine learning framework for small-sample scenarios, combining compositional characteristics, data augmentation, and Bayesian hyperparameter adjustment to improve prediction accuracy. Four regression models (ANN, SVR, decision tree, and LightGBM) were trained on proximate, ultimate, and petrographic features to predict coal calorific value. Among these, the LightGBM model achieved the highest predictive performance with a test (R² approximately 0.93) and the lowest error (RMSE approximately 0.19), outperforming the ANN (R² approximately 0.91) and other models. Fixed carbon (FC) and volatile matter (VM) were the key predictors of calorific value, aligning with domain knowledge and model interpretability. The improved data-driven approach reliably estimates coal energy content from small samples, enabling evidence-based geological modelling and better drilling decisions for increased exploration efficiency.