Machine Learned Interatomic Potentials for Prediction of Nanocatalysts Structures

Abstract

Catalysts are everywhere. They help accelerate and enable essential industrial processes towards success by being selective towards certain products and producing higher yields of said products at an agile manner. Nowadays, the relevance of catalysts is not only in industrial production, but also in the development of novel structures which are able to provide reaction pathways of relevant processes, such as the production of green hydrogen, or conversion of $CO_2$ into relevant products. For this, computational simulations are used as the first step in screening potential candidates that are able to provide higher yields or selectivity in heterogeneous catalysis reactions. However, these simulations are mainly done through the use of DFT, which requires a high computational cost and convergence time. Machine Learned Interatomic Potentials (MLIPs) have risen as complements for DFT simulations via training and learning from DFT energies and forces data to provide a platform for molecular dynamics simulations used to study the movement and behavior of atoms and molecules over time. In this research project, an active learning loop is engineered with the purpose of automating the workflow of training, using, and fine-tuning a MLIP (in this case, MACE) for its further use in catalysis energetics calculations.