Learning Transferable Neural Network Surrogates for Kohn-Sham Density Functional Theory

Density functional theory (DFT) is widely used to compute properties of many-body systems from first principles, and often serves as an intermediate step to computing forces on atomic nuclei in molecular dynamics (MD) simulations. However, DFT calculations are computationally intensive, and their cost becomes prohibitive for large systems or long MD simulations. Recently, efforts to circumvent expensive DFT calculations by learning system-specific neural network surrogates have been met with some success. Still, these surrogates require the generation of ab initio training data and long training times for every individual system and thermodynamic state. Here, we investigate whether this approach can be made more widely applicable by leveraging modern machine learning techniques such as transfer learning to produce surrogate models that transfer between systems and states, or adapt quickly without the need for extensive training. We develop first-of-their-kind models that are able to compute the local density of states directly from local descriptors of atomic positions across multiple temperatures, and demonstrate that the high-dimensional input descriptors may be compressed to a representation with lower information capacity without sacrificing prediction accuracy.