Accelerating Chemical Discovery with Reaction Simulations via Neural Network Potentials

Authors: Obadiah Smolenski, Shehtab Zaman, Huanyi Qin, Preston Kostioukov, Robin Cramer, Jason Zheng, Jonathan Zahavi

Field of Study: Computer Science

Faculty Mentors: Kenneth Chiu

Easel: 93

Timeslot: Morning

Abstract: Chemists often use simulations at the atomic scale when designing and analyzing new materials. However, traditional computational methods involving quantum mechanics (QM) are far too slow to use at a reasonable scale, taking days to run simulations on hundreds of atoms, while classical approximations miss vital events like bond breaking. As a result, graph neural networks (GNNs) have found great success as a compromise between accuracy and efficiency, because they can be trained on high-fidelity QM data without suffering from poor scaling with system size. In this work, using GNNs as chemical foundation models is demonstrated. With sufficiently diverse training data, they can simulate reactions with the accuracy of QM and comparable performance to classical approximations. In particular, this study sees that simulations run using GNNs on the combustion of methane and jet fuel, catalytic hydrogenation of CO2 on Pt, and electrolyte decomposition in Li-ion batteries correctly capture relevant reaction pathways.