(661c) Developing Physically Meaningful and Accurate Machine Learning Interatomic Potentials for Catalysis

Density Functional Theory (DFT) is the workhorse computational method to study catalytic reactions, which provides a good balance between accuracy and computational cost. Recent progress in machine learning interatomic potentials (MLIPs) can however change this evaluation by providing system specific solutions with DFT-level accuracy for orders of magnitude lower computational cost. I will detail my group’s efforts to develop physically meaningful and accurate MLIPs for various catalytically relevant systems including supported nanoparticles and solid-liquid interfaces. I will discuss our results on benchmarking various MLIPs and how MLIPs provide new avenues to make comparison with reliable experimental measurements to validate computational methods and protocols. Finally, I will highlight our developments to accelerate the training and execution of MLIPs as well as our recommendations on how to report MLIPs to make them findable, accessible, interoperable, and reproduceable to help accelerate the progress of the field.