Dresden 2026 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
TT: Fachverband Tiefe Temperaturen
TT 3: Correlated Electrons: Electronic Structure Calculations
TT 3.5: Talk
Monday, March 9, 2026, 10:30–10:45, HSZ/0103
Real-frequency DMFT for multi-orbital models with the neural network configuration interaction impurity solver — •Alexander Kowalski1, Philipp Hansmann2, Giorgio Sangiovanni1, and Adriana Pálffy1 — 1Institute for Theoretical Physics and Astrophysics, Universität Würzburg, 97074 Würzburg, Germany — 2Department of Physics, Friedrich-Alexander-Universität Erlangen-Nürnberg, 91058 Erlangen, Germany
The numerical solution of an auxiliary Anderson Impurity Model is usually the most resource-intensive part of the dynamical mean-field treatment of strongly correlated lattice models. Among the variety of approaches with different trade-offs that have been developed, the problem with exact diagonalization is that the exponential scaling of the Hilbert space constrains it to describing the infinite bath in terms of only a small finite number of sites. Selected Configuration Interaction (CI) methods try to achieve more favorable scaling by performing an a priori restriction of the basis used for diagonalization to try to capture only the most important subspace of the desired size. Neural networks have been shown to be remarkably suitable for performing the CI basis selection [1] and we have previously previewed the integration of such a CI algorithm as impurity solver in DMFT. Further developments since then have allowed us to do DMFT calculations with self-consistency directly on the real frequency axis for multi-orbital models with numbers of bath sites well beyond the reach of exact diagonalization.
[1] P. Bilous, L. Thirion, H. Menke, M. W. Haverkort, A. Pálffy, P. Hansmann, Phys. Rev. B 111, 035124 (2025)
Keywords: dynamical mean-field theory; configuration interaction; neural network