Machine Learning Predictions of Methane Storage in MOFs: Diverse Materials, Multiple Operating Conditions, and Reverse Models.

A machine learning (ML) model is developed for predicting useable methane (CH) capacities in metal-organic frameworks (MOFs). The model applies to a wide variety of MOFs, including those with and without open metal sites, and predicts capacities for multiple pressure swing conditions. Despite its wider applicability, the model requires only 5 measurable structural features as input, yet achieves accuracies that surpass less-general models. Application of the model to a database of more than a million hypothetical MOFs identified several hundred whose capacities surpass that of the benchmark MOF, UMCM-152. Guided by the computational predictions, one of the promising candidates, UMCM-153, was synthesized and demonstrated to achieve superior volumetric capacity for CH. Feature importance analyses reveal that pore volume and gravimetric surface area are the most important features for predicting CH capacity in MOFs. Finally, a reverse ML model is demonstrated. This model predicts the set of elementary MOF structural properties needed to achieve a desired CH capacity for a prescribed operating condition.