Robust Gaussian Process Regression method for efficient reaction pathway optimization: application to surface processes
Wei Fang,Yu-Cheng Zhu,Yi-Han Cheng,Yi-Ping Hao,Jeremy O. Richardson
2023-04-28
Abstract:Simulation of surface processes is a key part of computational chemistry that offers atomic-scale insights into mechanisms of heterogeneous catalysis, diffusion dynamics, as well as quantum tunneling phenomena. The most common theoretical approaches involve optimization of reaction pathways, including semiclassical tunneling pathways (called instantons). However, the computational effort can be demanding, especially for instanton optimizations with ab initio electronic structure. Recently, machine learning has been applied to accelerate reaction-pathway optimization, showing great potential for a wide range of applications. However, previous methods suffer from practical issues such as unfavorable scaling with respect to the size of the descriptor, and were mostly designed for reactions in the gas phase. We propose an improved framework based on Gaussian process regression for general transformed coordinates, which can alleviate the size problem. The descriptor combines internal and Cartesian coordinates, which improves the performance for modeling surface processes. We demonstrate with eleven instanton optimizations in three example systems that the new approach makes ab initio instanton optimization significantly cheaper, such that it becomes not much more expensive than a classical transition-state theory calculation.
Chemical Physics