Towards real dynamics in heterogeneous catalysis using machine learning interatomic potential simulations

Abstract

Heterogeneous catalysis plays a significant role in the modern chemical industry. Computational investigation has been an indispensable approach to reveal catalyst structures and catalytic reactions in the last few decades, where first-principles calculations are widely adopted. Despite the success in understanding catalysis at the atomic scale, the huge computational expense of such first-principles methods restrains the investigation to a simplified catalyst structure model and hinders the further exploration of complicated reaction networks. In tandem with the proliferation of computational catalysis studies and the advancement in computer science, Machine Learning (ML) as an emerging tool can take advantage of the accumulated data to emulate the output of ab initio methods with thousands of speedups, which makes high-throughput discoveries and large-scale simulations more efficient and more effective. In particular, Machine Learning Interatomic Potential (MLIP) has been a promising substitute of ab initio methods, which has a dramatically reduced computation cost while retaining a Density Functional Theory (DFT)-level accuracy. This thesis focuses on MLIP-based simulations towards realistic modelling of heterogeneous catalysis systems, which are performed either with a long time-scale or using a large-scale structure model. Chapter 1 discusses why realistic modelling is necessary for heterogeneous catalysis. Chapter 2 introduces the theoretical background and the computational approach, including some basic concepts of state-of-the-art ML techniques, which are frequently encountered in computational studies. Chapter 3 demonstrates an efficient framework I have been actively developing to automate the structure sampling and the MLIP training. Chapter 4 presents a study on CO oxidation on Pt(111) surface using Ab Initio Molecular Dynamics (AIMD) to reveal its dynamical behaviour. Chapter 5 proposes an algorithm that I use MLIPs to accelerate enhanced sampling methods for efficient Molecular Dynamics (MD)-based Free Energy Calculation (FEC). Chapter 6 studies the Pt surface oxidation with large-scale Grand Canonical Monte Carlo (GCMC) simulations using MLIP. Chapter 7 concludes this thesis and envisages how MLIP could further propel computational catalysis research in the near future.

Date of Award	Jul 2023
Original language	English
Awarding Institution	Queen's University Belfast
Sponsors	China Scholarship Council
Supervisor	Chunfei Wu (Supervisor) & Peijun Hu (Supervisor)