Machine learning-inspired process-based models for adsorption selectivity between CO₂ and CH₄ on coal

Adsorption selectivity of gases on porous substrates is a fundamental challenge in both materials science and applied engineering. Traditional models derived from single-component data and limited observations are insufficient to capture the integrated effects of temperature, pressure, and gas composition that govern multicomponent competitive sorption. We address this challenge by collecting then utilizing an ensemble data suite comprising 1017 entries for 43 coal samples under diverse conditions for CO₂/CH₄ competitive adsorption. This dataset was used to train four machine learning models to predict adsorption selectivity, among which XGBoost achieved the best performance. This ML-based model achieves a 14.6% improvement over traditional numerical models. SHapley Additive exPlanations (SHAP) analysis was performed to investigate the individual and interaction effects of key factors on adsorption selectivity. Pressure exhibits a non-monotonic three-stage influence on selectivity, governed sequentially by energetic, entropic, and pore-network effects. Coal rank, gas composition, and pressure exhibit pronounced interactions that reshape their individual contributions. Temperature suppresses CO₂ competitiveness, especially at low CO₂ fractions, due to its higher adsorption heat and enhanced thermal motion. This study establishes a data-driven framework integrating ML and adsorption physics for elucidating competitive adsorption in multicomponent gas systems, thereby advancing interdisciplinary inquiry across materials science, carbon sequestration, and natural gas development.