Abstract

Simulating chemical reactions is essential to developing a fundamental understanding of their mechanism and predicting experimental outcomes. Machine learned potentials (MLPs) offer an enticing approach to such simulations, enabling the efficient mapping between nuclear configurations and energies. Moreover, in contrast to classical force fields, they offer flexibility and systematic improvability. However, despite the development of Gaussian Approximation Potentials (GAPs) and high dimensional neural network potentials (NNPs) more than ten years ago, MLPs are still yet to find routine use for chemical reaction simulation. This slow uptake is likely due to the computational and time investment required to train reactive potentials for new systems, with only a handful of examples reported to date.

In this talk, I will present our team’s efforts to tackle these challenges by introducing efficient strategies to generate MLPs to study reaction mechanisms in the gas-phase and solution. Our work demonstrates that accurate potentials, achieving ab initio accuracy, can be generated employing only hundreds to a few thousand energy and gradient evaluations on a reference potential energy surface. I will also discuss the performance of different methods to obtain reactive MLPs for small to medium size reactions and discuss current limitations. I will finish by illustrating the power of the developed strategies in a diverse range of systems, including reactions in solution and ambimodal surfaces, as well as dynamical quantities, such as product ratios and free energies, for which expensive AIMD simulations would otherwise be needed. 

References

1. T. A. Young, T. Johnston-Wood, V. Deringer, F. Duarte. Chem. Sci., 2021,12, 10944.

2. T. A. Young, T. Johnston-Wood, H. Zhang, F. Duarte. Phys. Chem. Chem. Phys., 2022,24, 20820