(370g) Machine Learning Based Interatomic Potentials for Electrocatalysts

The design of next-generation functional materials and the quest to gain insights into dynamical processes at the atomic scale entail efficient and accurate atomistic simulations. The success of these simulations hinges on the accuracy, robustness, and transferability of the force fields (FFs) employed to describe interatomic interactions. One aspect of developing these FFs is to determine the set of variable parameters in their functional forms that closely reproduce available experimental or quantum mechanical data. In general, accurate FFs possess complex functional forms with numerous parameters, and call for sophisticated fitting strategies. Supervised machine learning methods, e.g., genetic algorithms in combination with local minimization methods (e.g., Simplex) provide an efficient way to scan the parameter space for such an optimization problem. Here, we employ such a machine learning approach to develop accurate FF for Ir-O and Ru-O systems; these Pt-group oxides are of immediate technological relevance in electrocatalysis, pH sensors, and photocatalytic splitting of water. For these oxide systems, we employ a Tersoff-like bond order potential to treat bond-directionality, and couple it with an electronegativity-equalization scheme (EEM) to accurately capture dynamic transfer for charges between various atomic species. We parameterize these FFs by training against extensive dataset obtained from first-principles calculations by employing genetic algorithms. Our newly developed Tersoff+EEM potential for these oxides accurately captures (a) structure, energetics, and elastic properties, (b) O adsorption kinetics, and (c) reaction pathways. These representative examples are used to highlight the effectiveness of evolutionary strategies in bridging the atomistic and electronic length scales.