Smart screening of hydrogen storage capacities in MOFs using a tailored machine learning

Metal-organic frameworks (MOFs) have emerged as promising candidates for solid-state hydrogen storage owing to their exceptional specific surface area, high pore volume, and chemically tunable structural properties. In this work, a diverse set of experimentally synthesized MOFs were evaluated to model and predict hydrogen storage capacity (wt%), using 4 key descriptors which are Brunauer–Emmett–Teller (BET) surface area, pore volume, operating pressure, and temperature. Correlation analysis revealed positive associations between BET surface area, pressure, and pore volume with storage capacity, and a negative association with temperature, consistent with physisorption mechanism. Six machine learning models were developed: support vector regression (SVR), artificial neural networks (ANN), random forest (RF), Gaussian process regression (GPR), gradient boosting (GB), and a Committee of Expert Systems (CES) integrating all base learners. While GB was the top-performing standalone model, the CES delivered the highest predictive fidelity (R² = 0.9958, MSE = 0.0094), as confirmed by parity plots and residual analysis. SHapley Additive exPlanations (SHAP) corroborated the statistical feature rankings, consistently identifying BET surface area and pressure as the most influential positive contributors in alignment with adsorption thermodynamics. Paired t-tests on root-mean-square error (RMSE) values confirmed statistically significant CES improvements over all individual models. The CES framework thus offers a data-efficient, accurate, and interpretable approach for rapid MOF screening, with straightforward adaptability to other porous materials and adsorption-based energy storage systems.