Large Machine-learned Potentials for Materials Design with Catalysis and Energy Applications

Abstract:

Designing multicomponent materials is challenging with density functional theory (DFT) because it computationally expensive to evaluate all the ways that elements may combine and react. It is also difficult to estimate free energy contributions to reactions and to locate reaction barriers with DFT. The Open Catalyst Project is developing machine learned potentials (MLP) to mitigate these challenges. These MLPs are trained on 100M+ DFT calculations spanning 55 different elements and 80+ adsorbates that are relevant in catalysis and energy applications. Nominally these models were trained to predict energy and forces, and from these one can derive reaction energies. We will show in this talk, however, that these models also...