Full Integration of Machine Learning Techniques with Nonadiabatic Molecular Dynamics Simulations

Nonadiabatic Molecular Dynamics (NAMD) simulations are widely used for exploring the time evolution of photoexcited molecular systems, which are able to access multiple potential energy surfaces during dynamics. Recent advancements in Machine Learning (ML) enabled the acceleration of NAMD using supervised learning methods to predict the necessary energies and forces for driving these simulations. This, in turn, can be exploited for improving statistical descriptions of photochemical reactions by increasing the number of trajectories being propagated. One significant challenge facing the adoption of ML techniques in NAMD simulations is the lack of clear protocols and software for training and deploying ML potentials, which hinders their widespread adoption. To address this challenge, I present a comprehensive end-to-end workflow for NAMD simulations with ML potentials, from geometry sampling and model training using active learning to the analysis of trajectories and properties using unsupervised learning, allowing the characterization of reaction mechanisms and photoproducts. This workflow is designed to be flexible and robust, applicable to different molecular systems and capable of handling large and complex datasets. It is implemented in two free and open source Python packages: MELTS for active learning and ULaMDyn for unsupervised learning, both integrated within the Newton-X software suite.