For the past 30 years, physics-based simulations using density functional theory (DFT) have powered our in silico understanding of the physical, chemical, and electronic properties of materials. These simulations are time- and compute-intensive, and can only simulate small collections of atoms (no more than 500) for no more than a few picoseconds. As one example of scale, these simulations use 30% of the Department of Energy’s Berkeley supercomputer facility.

‍

Recently, machine learning has been applied to accurately and quickly reproduce DFT-calculated physical properties of materials, but limited work has been done to machine learn the electronic structure of materials.

‍

RHOAR-Net seeks to construct a machine learning model to predict the electronic structure of materials. The project will pair these models with the physical property predictors to get a more complete understanding of materials.

‍

With the first foundation model to predict the electron density of diverse materials, we will be able to circumvent the most compute-intensive step of DFT, reducing the compute cost by orders of magnitude, and making it possible to study materials at unprecedented lengths and timescales.

‍