Machine Learning Interatomic Potentials for Reaction Pathway Prediction

Abstract

Reaction pathway prediction is crucial in computationally studying reaction thermodynam- ics, catalysis, transition state structures etc. Traditionally, this has been done using density functional theory (DFT) methods due to their accuracy. However, advancements in machine learning based methods may provide efficient computation at comparable accuracy. In this thesis, we examine the mathematical motivations and foundations of ACE and MACE ar- chitectures. Additionally, the reaction pathway was predicted for the cis-trans isomerism of but-2-ene using a MACE foundation model. The results reveal that MACE can construct accurate geometries of the transition state despite predicting incorrect energies. Addition- ally, finetuning the foundation model improved the energy prediction at a level comparable to DFT. This work shows that finetuning MACE foundation models can be a promising alternative for DFT calculations in certain scenarios.