(342bk) Deep Learning Molecular Force Field for Gaseous Adsorption in Metal-Organic Frameworks with Open-Metal Sites

Datar, A. - Presenter, The Ohio State University
Lin, L. C., The Ohio State University
Howe, J., Texas Tech University
Pandey, I., Texas Tech University
Yang, C. T., The Ohio State Universtity
Chen, C. C., Texas Tech University
Open-metal-site Metal−Organic Frameworks (MOFs) have exhibited promising performance in a variety of energy-related applications, particularly in gas separations. To facilitate the future design of new MOFs as well as to shed light on their atomic-level adsorption mechanism, molecular simulations play a critically important role. A prerequisite in molecular simulations is to properly describe the interactions between the adsorbates and the frameworks in order to ensure reliable predictions of relevant thermodynamic and transport properties. Although generic force fields have been shown to generate reasonable predictions, their failure to correctly capture the adsorption of CO2 in open-metal-site MOFs necessitates the need for developing accurate potentials. Machine learning (ML) has proven to be a powerful approach to make accurate predictions given unseen data. While it has been recently applied in the field of MOFs to accelerate the discovery of materials, its application to describe adsorbate-adsorbent interactions has not yet been explored.

In this study, we have for the first time developed a deep neural network (DNN) potential to model adsorbate-adsorbent interactions in open-metal-site MOFs and demonstrated its potential in adsorption simulations. Our DNN model is based on physical interactions-guided features that utilize terms in the classic force fields and the pair-distance. The training and test sets were generated using state-of-the-art density functional theory (DFT) simulations. Both non-polar (e.g., CO2) and polar (e.g., H2O and CO) molecules are considered and studied in this work. Using Mg-MOF-74 as a case study, with as few as 1000 adsorbed configurations with their reference energies computed by DFT, sophisticated DNN models can be established and are capable of accurately describing the adsorption energy of gases. We have also demonstrated the promising computational efficiency of such DNN potential in the calculations of adsorption properties such as Henry’s constants in porous materials. The approach presented herein is expected to be also applicable for MOFs without open-metal sites or porous materials in general. Overall, this study is anticipated to pave the way towards the future development of highly accurate molecular potential for use in molecular simulations to accelerate materials discoveries.