Neural Network potentials are developed which accurately make and break bonds for use in molecular simulations. Large, condensed phased, and extended systems remain a challenge for theoretical studies due to the compromise between accuracy and computational cost in calculations. Machine learning methods provide a new approach to solve this trade off by leveraging large datasets to train on highly accurate calculations on small molecules and then extending learned features to predict larger systems. Specifically, neural network potentials have shown great promise to achieve high accuracy calculations with low computational cost. In this project, we introduce a novel method to combat the accuracy versus cost battle in computational chemistry.<br/>We designed a training algorithm to generate neural network potentials to predict molecular energies of specific systems of interest; this method aims to use the minimum amount of data of smaller/sub systems to train neural network potentials to the accuracy level of advanced <i>ab initio </i>methods. Starting with a neural network potential trained at the density functional theory (DFT) level, we generate homolytic carbon-carbon bond dissociation data of small size alkanes with DFT energies to train the potentials to accurately predict bond dissociation at the DFT level. Then, using transfer learning, we retrained the neural network potential to complete active space second-order perturbation theory (CASPT2) level of accuracy. We demonstrate that the neural network potential only requires bond dissociation data of a few small alkanes to accurately predict bond dissociation energy in larger alkanes. This training algorithm can be further applied to any type of bond or any level of theory.

Machine learning potentials for bond breaking: Training CASPT2 level neural network potentials for carbon carbon dissociation in alkanes

ACS Spring Meeting 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.