Deep neural networks to accelerate and reproduce DFT

Databases such as the Open Quantum Materials Database and the Materials Project contain the results of density functional theory (DFT) calculations for hundreds of thousands of materials structures. Data at this scale allows us to leverage machine learning models to accelerate DFT computations, or to completely replace them. In this talk we present our recent progress in deep neural network models that can accelerate or reproduce DFT predictions, including energies, band gaps, and electron densities. When training for multiple targets, these networks also generate reduced-dimensional latent space representations that may act as materials fingerprints.