Incorporating prior physical knowledge into the development of learning algorithms is critical for incubating transferable and white-box models. Bayesian learning, in particular, Gaussian processes (GPs), provides a formalism in which the prior knowledge about the system can be included as the prior mean functions, and the predictions – the posterior means – are refined by learning. In this talk, I will demonstrate how to merge prior understanding (e.g., a classical force field approximation, a low-level semi-empirical approximation) of molecular potential energy surfaces (PESs) with GPs to learn and explore the PESs simultaneously. I will use geometry optimization as an example of PES explorations and show that our physics-inspired GPs are able to accelerate molecular geometry optimization significantly. Our GPs can be trained in a multi-level fashion, in which a pre-learned PES serves as the prior mean functions in the next-level GP for high-accuracy molecular geometries. Noticeable features of these learning models include the elimination of a pre-defined training set and on-the-fly learning with a limited number of data points.

Physics inspired Bayesian learning for potential energy surface exploration

Where thousands of chemistry professionals meet to share ideas and advance scientific and technical knowledge. Spring 2023 abstracts, posters, and presentations will launch at the end of March.

ACS Spring 2023

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.