26 July 2026 to 1 August 2026
University of Maryland, College Park
US/Eastern timezone

Machine-Learned Density of States for the Doped Hubbard Model

28 Jul 2026, 16:30
20m
Thurgood Marshall (Adele H. Stamp Student Union)

Thurgood Marshall

Adele H. Stamp Student Union

3972 Campus Dr, College Park, MD 20742
Contributed talk Algorithms and artificial intelligence Algorithms and artificial intelligence

Speaker

Felicitas Freche

Description

In this work, we extend the normalizing-flow-based generalized density-of-states (NF-gDoS) method to the doped Hubbard model. The Hubbard model is known to exhibit a sign problem in the presence of a chemical potential. The NF-gDoS framework is attractive because it reformulates the original complex-weight sampling problem by separating the complex phase from the Boltzmann distribution and encoding it instead in a one-dimensional oscillatory integral.
Recent studies have demonstrated that the NF-gDoS method can successfully reconstruct the partition function in 1+1D scalar field theory, correctly reproducing the Lee–Yang zeros. In a separate study of 1+1D U(1) gauge theory with a $\theta$-term, it was shown that the density of states can be qualitatively reconstructed, highlighting the importance of expressive normalizing flows.
Building on these results, we apply expressive conditional-normalizing-flow (CNF) architectures to the doped Hubbard model and systematically benchmark the resulting density of states against exact results for small system sizes. Our study provides a first assessment of the applicability of CNF-gDoS to a strongly correlated fermionic system. We identify key challenges, including limitations related to model expressivity and stability, and outline possible directions for improving the robustness of the approach in more demanding regimes.

Author

Co-authors

Lena Funcke (University of Bonn) Janik Kreit (University of Bonn) Dominic Schuh (University of Bonn) Simran Singh (HISKP, University of Bonn)

Presentation materials