(306h) Discovering Equations for Acid/Base Interactions Using Symbolic Regression

Machine learning algorithms extract patterns from large datasets to generate predictions. Some tools lead to interpretable models, and some learning algorithms are informed by physics, but generally, these are not easily related to the equation-based theories and derivations in the literature. Symbolic regression is a tool for generating equations that match data, and our recent work has augmented symbolic regression algorithms with logical reasoning capabilities.

In this talk, we describe how we apply these methods to the task of force field fitting. We show how symbolic regression can discover short, closed-form expressions for molecular interactions that don't currently have established force fields, and how this can enable accelerated simulations for molecules interacting with catalytic materials, especially Brønsted acidic zeolites.