Learning the electronic density of states in condensed matter

The electronic density of states (DOS) describes the energy levels accessible to electrons in a quasi-particle picture. It is essential to interpret experimental observables such as heat capacity, or optical absorption. In this work, we present a machine-learning (ML) model to learn the DOS in different classes of materials, discuss the challenges of predicting it, provide insights on more complex systems and quantify finite-temperature effects. We introduce an atom-centered model for the electronic DOS where we expand the total DOS of a structure into a sum of contributions from its atomic environments. This ML model provides quantitative predictions of the DOS and its derived quantities like the band energy in our validation sets. We successfully apply the model to predict a hybrid-DFT quality DOS of large amorphous Silicon structures covering a wide range of pressures. We also employ the model to account for the electronic contributions to the heat capacity in metallic systems. This approach demonstrates the impact of a universal model describing structural and electronic properties inexpensively and its ability to enable more accurate and predictive materials modeling and design.