AI for Materials Discovery II

Equivariant neural networks (ENNs) trained on density functional theory (DFT) calculations have shown strong utility in property prediction and simulation tasks. However, DFT-based datasets suffer from both high compute cost to generate, as well as systemic errors caused by the intrinsic modeling assumptions of DFT. Certain types of experimental measurements can be both cheaper and faster to perform relative to an ab-initio calculation of the physical observable that is being measured. Moreover, we may regard the experimental measurement as closer to 'ground truth' than a computed approximation. Various works have individually used each of these two types of data for prediction tasks, but few empirical investigations have been performed on potential performance gains when using both in conjunction. In this work, we present an architecture and workflow for training an ENN on both ab-initio calculations and experimental measurements for the purposes of making direct predictions of material properties of interest. Additionally, we provide analyses of out-of-domain prediction ability, scaling laws, and performance comparisons to single-modal models that utilize only one form of data.