Machine Learning-Based Potential Energy Surfaces for Molecular Simulation

Markus Meuwly

In this contribution I will discuss recent advances in conceiving, using, and refining molecular potential energy surfaces for atomistic simulations in the gas phase and in solution. The focus is on high-accuracy models for which direct comparison with experiments is possible to delineate the range of validity of such PESs. Applications range from infrared spectroscopy to the dynamics of bimolecular reactions and the characterization of product state distributions, all of which are also amenable to experimental investigations. A second topic that will be briefly discussed concerns uncertainty quantification and validation of the trained PESs from a numerical perspective.