Coupling Molecular Simulation and Machine Learning for CO2 Adsorption Prediction on Defective and N/O/S-Doped Graphene Models: Insights into Feature Importance and Synergistic Effects

Abstract CO2 emissions were recognized to exacerbate climate change. Porous carbon-based adsorption was identified as a promising industrial sequestration strategy. However, complex structure-performance relationships could not be fully elucidated via traditional experiments, and classical models failed to predict nonlinear adsorption. In this study, a hybrid framework integrating density functional theory (DFT), grand canonical Monte Carlo (GCMC) simulation, and machine learning was developed to predict CO2 adsorption on pristine, defective, and N/O/S-doped graphene models. Atomic models (pristine, defective, N/O/S-doped graphene) were constructed and optimized. CO2 isotherms were calculated (273–323 K, 0–10 bar) to obtain the maximum monolayer adsorption capacity. Five ML models were trained, tested, verified, and compared. Principal component analysis was applied for feature importance. Results revealed heteroatom doping accounted for >85% capacity variation (internal O, external S, internal S as top features), defects imposed secondary effects, all models achieved R2 > 0.98, and random forest (RF) exhibited optimal performance (R2 = 0.9834, MSE < 0.030). It was proved that the RF model captured nonlinear correlations and confirmed O–S codoping synergy. This work provides atomic-level insights into defect and doping effects and offers a data-driven approach for the rational design of carbon-based CO2 adsorbents.