ES22-Modine

Author: Modine, N.A., Sandia National Laboratories, Albuquerque, New Mexico

Title: A machine learning surrogate for density functional theory based on the local density of states

Abstract: We present a workflow based on machine learning (ML) that can reproduce Kohn-Sham density functional theory (DFT) total energies to within chemical accuracy (< 1 kcal/mol) with a computational cost that scales linearly with system size. This workflow uses a deep neural network to predict the local density of states (LDOS) as a function of the nearby arrangement of atoms, which is encoded using spectral neighbor analysis potential (SNAP) descriptors. From the LDOS, spatially-resolved, energy-resolved, and integrated quantities can be calculated, including the DFT total energy and forces, and we will discuss implementation of these calculations. Once the ML model has been trained on the LDOS for cells where DFT calculations are practical, the model can be accurately and efficiently applied to much larger systems. Compared to DFT, the only new approximation in this approach is the determination of the LDOS using a local ML model rather than by solving the Kohn-Sham equations. In particular, the treatment of charge transfer through the determination of a global Fermi level and the incorporation of long-ranged electrostatic contributions to the energy are exactly the same as DFT. We will propose application of this approach to several problems that are very challenging for conventional interatomic potentials including: (1) charged defects in semiconductors and insulators, (2) systems where the effects of electronic temperature need to be represented, and (3) metal-insulator transitions induced by structural changes.

Sandia National Laboratories is a multimission laboratory managed and operated by National Technology & Engineering Solutions of Sandia, LLC, a wholly owned subsidiary of Honeywell International Inc., for the U.S. DOE's National Nuclear Security Administration under contract DE-NA0003525. The Center for Advanced Systems Understanding (CASUS) is financed by the German Federal Ministry of Education and Research (BMBF) and by the Saxon Ministry for Science, Art, and Tourism (SMWK) with tax funds on the basis of the budget approved by the Saxon State Parliament.

Other authors: Fiedler, Lenz 2,3; Vogel, D.J. 1; Thompson, A.P. 1; Ellis, J.A. 4, Stephens, J.A., Popoola, G.A., Cangi, Attila 2,3, and Rajamanickam, S. 1 1 Sandia National Laboratories, Albuquerque, New Mexico 87185, USA 2 Center for Advanced Systems Understanding (CASUS), D-02826 Görlitz, Germany 3 Helmholtz-Zentrum Dresden-Rossendorf, D-01328 Dresden, Germany 4 Oak Ridge National Laboratory, Oak Ridge, Tennessee 27830, USA