Rapid prediction of single-site adsorbate probability distributions in metal–organic frameworks using graph neural networks

Abstract

Metal–organic frameworks (MOFs) are porous crystalline materials assembled from inorganic nodes and organic linkers. These materials have garnered significant interest for gas separation and storage applications, particularly because of their porosity and their tunability due to their massive design space. However, navigating such a massive design space poses significant challenges. Atomistic simulation techniques have been applied to accelerate discovery and design of MOFs for various applications. A key property obtained from these simulations is the adsorbate probability distribution (APD). An APD maps the probability of finding an adsorbate molecule in the pore of a MOF at a given temperature and pressure, whose maxima correspond to free energy minima (i.e., binding sites). While APDs and binding sites are not easily accessible experimentally, their generation via simulation is tractable. However, high-throughput generation of APDs still requires long simulation times to converge. A machine learning (ML) model to predict APDs would enable the use of this property in data-driven pipelines to identify high performing materials or binding sites. To date, nobody has attempted to apply ML to the prediction of APDs or binding sites of MOFs. In this work, we present DeepAPD – an ML model which predicts APDs at a given temperature and pressure. As an initial proof of concept, the model has been trained on simple spherical adsorbates such as CH₄ and Xe. DeepAPD was found to generate APDs of MOFs at a speedup factor of >10⁵ in comparison to GCMC. An in-depth discussion of training strategies and dataset size/composition on model performance is presented. It was found that the APDs obtained by ML were sufficiently accurate to get a reliable estimation of binding sites in MOFs, particularly binding sites which have high probability. Finally, the transferability of the ML models was investigated by evaluating the performance of the GNN model on a dataset of experimentally characterized MOFs. We have also implemented the DeepAPD inference code into our binding site identification algorithm to facilitate an end-to-end MOF to binding site prediction. Future work will extend these models to more complex guests such as CO₂, N₂, and H₂O.