It’s no secret that machine learning (ML) is sweeping across the quantum chemistry world. What ML models offer in terms of speed and accuracy, however, is not always matched by their explanatory value. In this work I will present my work building an ML model that, in addition to being very accurate, gives insight into the “near-sightedness” of electron correlation. This ML model uses a new, physics-informed class of algorithm called Euclidean Neural Networks to predict the electron density of molecular systems. This electron density can then be used with the Hellmann-Feynman Theorem to compute forces for very large systems. I will show that this scheme allows two important scientific advances. First, it allows an ML Quantum Chemistry model to extrapolate. The model can be trained on small-scale calculations, but still make accurate predictions on large systems. And second, this scheme allows us to delineate between the “classical” and “quantum” contributions to molecular dynamics. I will show that by systematically varying the size of a ML training set of water clusters, one can establish an effective “radius of correlation”. This suggests a promising path forward for physics-informed machine learning methods.

What can machine learning and the Hellmann-Feynman Theorem teach us about the limits of electron correlation?

ACS Fall 2021

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