Physics Informed Neural Nets for Prediction of KS Potentials

Kohn-Sham density functional theory is one of the most successful electronic structure methods for molecules and materials, and density-to-potential inversions can provide insights into the exact formalism underlying this approach. This work looks to circumvent normal inversion schemes by employing Physics Informed Neural Nets (PINNs) in their place. PINNs help to improve predictive transferability and reduce the requisite amount of data to properly train a neural network. Implementations of a convolutional PINN and its application to exactly solvable models, such as soft-Coulomb systems, will be presented. Extensions of the network into ensemble density functional theory and realistic systems will be discussed.