Recent advances in machine learning for electronic excited state molecular dynamics simulations

Abstract

Machine learning has proven useful in countless different areas over the past years, including theoretical and computational chemistry, where various issues can be addressed by means of machine learning methods. Some of these involve electronic excited-state calculations, such as those performed in nonadiabatic molecular dynamics simulations. Here, we review the current literature highlighting recent developments and advances regarding the application of machine learning to computer simulations of molecular dynamics involving electronically excited states.

CASNet: Learning Complete Active Space Orbitals using Message Passing Neural Networks

“…In recent years, machine learning (ML) has emerged as a promising tool for mitigating the high computational expenses typically associated with quantum chemistry [11]. Particularly secondary outputs, for example energies and forces, have been proven to be a suitable learning target for ML, both in the ground state [12][13][14][15] and excited states [10,16,17]. However, these methods require separately trained models for most properties, resulting in a certain level of redundancy and potential inconsistencies.…”