Combining crystal graphs and domain knowledge in machine learning to predict metal-organic frameworks performance in methane adsorption

Abstract

A novel method capable of efficiently predicting the methane adsorption uptakes by metal-organic frameworks (MOFs) is proposed, which is built on the crystal graph convolutional neural network algorithm. Our method considers certain key physical features of MOFs along with information on secondary building units. Because the general force field fails to reproduce methane–MOF interactions involving the open metal site (OMS) adequately, the new force field parameters for methane are refined specifically for the OMS MOFs using a generic algorithm, which is particularly useful in constructing a more accurate dataset via the high-throughput grand canonical Monte Carlo. Extensive analyses demonstrate that the Pearson correlation coefficient and mean average error can reach as high as 0.973 and 9.98 cm3/cm3 for the newly developed algorithm, respectively. Furthermore, this model acquires information from the adsorption volume as embedding representations. This allows for the application of transfer learning by fine-tuning a pretrained graph network for the different properties. Finally, we predicted the methane adsorption volumes by all the MOFs in a 137953 hypothetical database at 298 K and 65 bar within hours. It is believed that this new method could be a useful tool in the early stages of high-throughput virtual screening of novel porous materials for gas adsorption or separation.