Optimized Machine Learning Potential Reconstruction for Enhanced Catalysis Research

In this work, we qualify the necessity of an improved workflow for constructing accurate, efficient machine learning (ML) models for computational catalysis research. Computational modeling of heterogeneous catalysts is largely dominated by ab initio methods such as density functional theory (DFT). DFT can be used to calculate stable and metastable catalyst structures which are important for investigating catalytic properties. Global optimization algorithms such as the genetic algorithm (GA) can apply DFT to explore the potential energy surface of catalytic systems and identify structures of interest, but the cost of DFT rapidly scales with the number of atoms being simulated. This computational cost constrains global optimization studies to catalytic systems of a few hundred atoms or less. Algorithms employing DFT combined with a surrogate ML model have been shown to be 2-3 orders of magnitude less expensive than their pure DFT counterparts, so the application of ML models to increase the speed and size of simulations is of great interest for researchers. Unfortunately, constructing ML models with accuracy close to that of DFT is an involved process requiring familiarity with the underlying mathematics of the model and a priori knowledge of the interactions present in the system. The workflow we are presenting can rapidly build models for new systems. We will prove this by creating ML surrogate models for two gas-phase nanocluster systems and one supported nanocluster system. We hope to show that this workflow reduces the barrier to using ML models for improved computational catalysis research.tsigo