Background

Metal-organic frameworks (MOFs) have been recently studied as promising hydrogen storage adsorbents. However, large-scale screening by experiments is infeasible. Machine learning and Grand Canonical Monte Carlo (GCMC) have been widely applied in the screening of MOFs to greatly improve efficiency.

Methods

GCMC was applied to 418 MOFs to calculate structural and performance parameters, and the gradient boosted regression (GBR) machine learning algorithm was used to predict the hydrogen adsorption capacity at cryogenic temperatures and high pressures from room temperatures and low pressures, and the database was extended to 8024 MOFs adsorbents based on the established ML model. The influencing factors affecting hydrogen adsorption capacity were comprehensively evaluated, and four representative MOFs were randomly selected for hydrogen adsorption isotherm testing, and the accuracy of ML was verified by comparing the GCMC simulation results with those predicted by ML.

Significant findings

The accuracy of the ML model is acceptable, with the coefficient of determination (R²) of the train all exceeding 0.9 and the test R² all exceeding 0.85. The MOFs exhibiting high hydrogen adsorption capacity by low-pressure adsorption have low pore size, Henry's constants, and density, and the metal-based is Al-based and Zn-based, while the surface area dominates at high-pressure adsorption. The experimentally tested MOFs with hydrogen adsorption capacity compared with GCMC simulation and ML models have an error of 1.01% in the low-pressure region and 1.49% in the high-pressure region.