Data-driven explainable artificial intelligence models for predicting CO₂ uptake in metal-organic frameworks

This study presents a comprehensive data-driven framework for predicting and interpreting CO₂ uptake in metal-organic frameworks (MOFs) using advanced machine learning (ML) and explainable artificial intelligence (XAI) techniques. A curated dataset of 223 experimentally reported MOF samples, characterized by surface area (SA), pore volume (PV), temperature (T), and pressure (P), was used to train four supervised ML models: Artificial Neural Network (ANN), Random Forest (RF), AdaBoost (AdB), and Gradient Boosting (GB). Among the evaluated models, the GB model achieved the highest predictive accuracy, yielding low training errors (MSE = 0.398, RMSE = 0.63, MAE = 0.211, RRMSE = 0.49, R² = 0.99) and maintaining strong generalization on the test set with an MSE, RMSE, MAE, RRMSE, R² as 3.752, 1.937, 1.041, 1.495 and 0.982. To ensure model transparency, SHapley Additive exPlanations (SHAP) were employed, revealing "P" and "PV" as the most influential features governing CO₂ adsorption, while "T" showed minimal impact. The interpretability analysis also highlighted nonlinear feature interactions and saturation effects associated with surface area and pore characteristics. Finally, a graphical user interface (GUI) was developed using the optimized GB model, enabling real-time CO₂ uptake prediction based on user-defined inputs. This integrated prediction-interpretation-deployment framework provides a practical and scientifically transparent tool for rapid MOF screening, supporting data-driven advancements in carbon capture and sustainable materials design.