DMFT-QE Symposium: March 30th

Name: DMFT-QE Symposium: March 30th
Start: 2026-03-30T09:30:00-04:00
End: 2026-03-30T11:00:00-04:00
Location: Virtual

Monday March 30, 2026
9:30 - 11:00 a.m.

Add to Calendar

Virtual

Invitation Only

Talk 1:

Neural-Network Quantum Impurity Solvers for Accelerating DMFT

Agnes Valenti, Flatiron Institute, CCQ

Quantum impurity solvers are the computational bottleneck of quantum embedding approaches to correlated materials, such as dynamical mean-field theory (DMFT). In this talk, I will introduce a set of neural-network impurity solvers (“GNet”) that, by initializing controlled calculations, dramatically accelerate DMFT while preserving accuracy.

In particular, GNet is trained to learn the impurity mapping from hybridization functions and local interaction to Green’s functions. I will explain that training on synthetic, material-agnostic data allows to achieve quantitative accuracy for both model systems and real materials. The resulting fast solvers for single-and multi-orbital models are benchmarked against CTSEG and CTHYB, showing that the method reproduces the Mott transition, multi-orbital phase diagrams of Hubbard-Kanamori models, and the electronic properties of SrVO$_3$ and SrMnO$_3$.

Talk 2:

Machine-Learned Self-Energies for Accelerating DFT+DMFT: Application to Iron at Earth’s Core Conditions

Li Zhi, Rutgers

Charge self-consistent DFT+DMFT provides a powerful quantum-embedding framework for finite-temperature correlated materials, but repeated impurity solves make large-scale thermodynamic sampling prohibitively expensive. In this talk, I will present a physics-informed machine-learning approach that accelerates DFT+DMFT by predicting the local self-energy and Fermi level from atomic environments. Rather than learning the Matsubara self-energy as an unconstrained complex function, we learn its high-frequency limit together with a compact set of Legendre coefficients using E(3)-equivariant graph neural networks. These predictions provide a warm start for the DMFT self-consistency loop and reduce the number of iterations by a factor of 2-4 across Fe, FeO, and NiO. I will then show how this acceleration enables high-throughput DFT+DMFT calculations for Fe at Earth’s core conditions, training of a neural-network interatomic potential from DFT+DMFT energies and forces, and two-phase coexistence simulations of melting. The resulting melting temperature of Fe at 330 GPa is 6225 K, in good agreement with recent experimental constraints, and illustrates how machine learning can make quantum-embedding calculations practical at much larger scales.