Multi-fidelity Graph Networks for Machine Learning the Experimental Properties of Ordered and Disordered Materials

Predicting the properties of a material from the arrangement of its atoms is a fundamental goal in materials science. In recent years, machine learning (ML) on ab initio calculations has emerged as a new paradigm to provide rapid predictions of materials properties across vast chemical spaces. However, the performances of ML models are determined by the quantity and quality of data, which tend to be inversely correlated with each other. Here, we develop multi-fidelity materials graph networks (MFGNet) to transcend this trade-off to achieve accurate predictions of the experimental band gaps of ordered and disordered materials to within 0.3-0.5 eV. We show that the inclusion of low-fidelity Perdew-Burke-Ernzerhof band gaps significantly enhances the resolution of latent structural features in materials graph representations, leading to 22-45% decrease in the mean absolute errors of high-fidelity computed and experimental band gap predictions with an order of magnitude smaller data sizes. Further, MFGNet models can be readily extended to predict the band gaps of disordered crystals to excellent agreement with experiments, addressing a major gap in the computational prediction of materials properties.