Learning Simple Chemical Heuristics to Model and Discover Materials

Abstract

Computational approaches have long played an important role in the field of materials science, driving both the scientific study of materials’ fundamental properties and the design of materials for technological applications. Currently, mainstream methods in computational materials science typically rely on either first-principles calculations or deep learning models. In this thesis, we take a different direction by developing remarkably simple data-driven models for predicting fundamental properties of materials, including electronic topology, metallicity, and band gap. These models take the form of highly interpretable chemical heuristics. A key finding of this work is the surprising result that electronic topology diagnosis – often regarded as a highly complex task – can, in fact, be performed heuristically using a simple and intuitive model. We further integrate this model into a workflow for discovering new topological materials. Altogether, this work revisits the classic idea of chemical heuristics through a modern data-driven lens, shedding new light on fundamental problems in materials science.