Parts | Days | Selection | Search | Updates | Downloads | Help
DS: Fachverband Dünne Schichten
DS 4: Frontiers of Electronic-Structure Theory: Focus on the Interface Challenge I (joint session O/CPP/DS/TT)
DS 4.2: Talk
Monday, April 1, 2019, 11:00–11:15, H9
The Teacher and the Student: Exchange-Correlation Energy Densities from Quantum Chemistry and Machine-Learning — •Johannes T. Margraf, Christian Kunkel, and Karsten Reuter — Chair for Theoretical Chemistry, Technische Universität München, Germany
(Semi-)local density functional approximations (DFAs) are the workhorse electronic structure methods in condensed matter theory and surface science. Central to defining such DFAs is the exchange-correlation energy density єxc, a spatial function that yields the exchange-correlation energy Exc upon integration.
Unlike Exc, єxc is not uniquely defined. Indeed, there are infinitely many functions that integrate to the correct Exc for a given electron density ρ. The challenge for constructing a useful DFA is to find a systematic connection between ρ and єxc. While several empirical and rigorous approaches to this problem are known, there has been little innovation with respect to the fundamental functional forms of DFAs in recent years.
Herein, we discuss two less explored routes to constructing DFAs. Specifically, a recipe for deriving єxc directly from many-body wavefunctions is compared to a machine learning (ML) approach that infers the optimal єxc for a given functional form. We find that local DFAs based on the many-body єxc are not transferrable between systems because the underlying energy densities are inherently non-local. In contrast, the ML єxc is by construction as local as possible. The extension of both approaches to non-local DFAs will be discussed.