Publications Repository - Helmholtz-Zentrum Dresden-Rossendorf

I will talk about two recent efforts to apply advanced machine learning methods to the electronic structure problem at the density functional theory (DFT) level. First, I will present a machine learning approach based on physics-informed neural networks and neural operators for inverting the Kohn-Sham equations for the exchange-correlation (XC) potential; neural networks provide a new way to perform DFT inversions at scale by learning the mapping from density to potential [1]. Second, I will present a very recent development in which we use neural operators to predict the electron dynamics of systems driven by a laser field. This approach complements conventional numerical solvers and has the potential to enable real-time, on-the-fly modeling of laser-irradiated molecules and materials with varying experimental parameters [2]. Both methods are illustrated on a conceptual level using one-dimensional models of diatomic molecules, but the approach can be readily applied to realistic systems in three dimensions.

[1] V. Martinetto, K. Shah, A. Cangi, A. Pribram-Jones, Mach. Learn.: Sci. Technol. 5, 015050 (2024).
[2] K. Shah, P. Stiller, N. Hoffmann, A. Cangi, arXiv:2210.12522 (2022).