Dominant mechanism and threshold behavior of spontaneous combustion in coal gangue hills: A multi-field coupled and interpretable machine learning approach

The spontaneous combustion of coal gangue hills is a typical multi-field coupled process driven by airflow seepage, oxidation reactions, and heat accumulation. However, under multi-factor coupling conditions, the dominant controlling factors governing spontaneous combustion risk and their regulatory mechanisms remain insufficiently quantified, which hinders accurate prediction and effective engineering prevention of spontaneous combustion hazards in coal gangue hills. To address this issue, this study proposes a comprehensive CFD-OD-ML integrated framework, systematically investigates the coupled evolution characteristics of multi-physical fields within coal gangue hills, and quantitatively identifies the key controlling factors of spontaneous combustion. High-precision prediction of temperature evolution and combustion risk is achieved through a data-driven modeling approach. The results indicate that the spontaneous combustion risk is primarily governed by the competitive interaction between oxidation heat release and convective heat dissipation. Among the influencing parameters, porosity plays a dominant role, followed by carbon content and ambient temperature, whereas the effect of wind speed is comparatively weaker. Further analysis reveals a critical porosity threshold of approximately 0.35. Below this threshold, limited airflow suppresses convective heat dissipation, leading to heat accumulation and an increased risk of spontaneous combustion. Above the threshold, enhanced permeability strengthens convective heat transfer, thereby mitigating further temperature rise. The GP-LightGBM spontaneous combustion risk prediction model of coal gangue hills was constructed, and more than 96% of the prediction error was controlled within ± 20%. SHAP-based interpretability analysis further quantitatively confirms the dominant contribution of porosity and elucidates the nonlinear response and threshold effects of key variables. In addition, a GUI-based rapid prediction tool was developed to enable integrated application of CFD simulation outputs, orthogonal experimental design results, and machine learning predictions. Overall, this study provides new insights into the mechanism understanding, risk assessment, and engineering prevention strategies for spontaneous combustion in coal gangue hills.