GW+EDMFT Method Unifies Local and Nonlocal Quantum Correlations in Correlated Materials
Researchers at the Flatiron Institute, working with collaborators at the University of Fribourg in Switzerland and Université Paris-Saclay in France, have built a fully self-consistent GW+EDMFT method that models both local and nonlocal quantum correlations in materials.
The Tech TL;DR:
- The new GW+EDMFT implementation treats strong local interactions and longer-range screening in one self-consistent loop, instead of stacking approximations.
- Interpolative separable density fitting (ISDF) compresses two-particle correlation functions, cutting the computational cost that made such simulations impractical.
- On strontium manganese oxide the method reproduces a Mott-insulating state in quantitative agreement with experiment; on lanthanum nickel oxide it yields a U(0) of 9-10 electron volts with a J of 0.7-1 electron volt.
Why overscreening broke the older methods
Strongly correlated quantum materials are where plain density functional theory falls over: electrons interact strongly enough that simpler, perturbative approaches underestimate the interactions and systematically mispredict whether a material conducts.
Constrained-RPA, the standard workhorse for feeding interaction strengths into DFT+DMFT calculations, overestimates screening — a failure the team calls the “overscreening inherent to constrained-RPA approaches.” For modelers running these calculations on HPC clusters, that overscreening translated directly into wrong answers about metal-insulator transitions.
The fix in GW+EDMFT is architectural. The GW approximation handles nonlocal correlations and long-range screening, while extended dynamical mean-field theory treats strong local interactions.
Running both to full self-consistency — the Green’s function and the dynamically screened interaction — includes a “well-defined double-counting correction” and a “self-consistently determined impurity interaction,” per the paper, “Electronic correlations and dynamical screening with ab initio quantum embedding,” by Chia-Nan Yeh, Francesco Petocchi, Alexander Hampel, Philipp Werner, Olivier Parcollet, Antoine Georges and Miguel A. Morales.
ISDF compression made the arithmetic tractable
The bottleneck was never the physics formalism — it was the data volume. Two-particle correlation functions scale brutally with system size, and achieving self-consistency in both the Green’s function and the screened interaction previously made realistic materials infeasible. ISDF compresses those functions so the two channels can be converged simultaneously.
The payoff shows up in the numbers. Modeling lanthanum nickel oxide, the method produced a screened interaction that interpolates smoothly between the high-energy behavior predicted by self-consistent GW and the low-energy behavior from G0W0, with the crossover falling between 5 and 10 electron volts.
The effective local interaction U(0) lands at 9-10 electron volts — slightly above the values used in prior DFT+DMFT studies — with a Hund’s coupling J of 0.7-1 electron volt. Constrained-RPA only approaches that U(0) when screening within the nickel-d and oxygen-p orbitals is removed entirely by hand.
The interaction parameters a materials engineer would feed into a downstream DFT+DMFT run therefore come out of the calculation itself rather than a correction:
# LaNiO3 impurity parameters, from self-consistent GW+EDMFT
[hamiltonian]
U0 = 9.5 # eV, fully screened local interaction (range 9-10)
J = 0.85 # eV, Hund's coupling (range 0.7-1)
# constrained-RPA required zeroing d-p screening to match U0;
# GW+EDMFT reaches it from first principles.
SrMnO3 as the validation case
Strontium manganese oxide is the harder test: a Mott insulator where perturbative approaches fail outright. GW+EDMFT stabilizes the Mott-insulating state and does so “in quantitative agreement with experiment.”
The team found the fully screened interaction in the insulating phase stays close to the local effective interaction because impurity polarization is strongly suppressed — vertex corrections crush charge fluctuations and yield a much larger, less structured effective local interaction than constrained-RPA predicts.
The contrast with the metallic LaNiO3 is the point. Local EDMFT corrections there amplify charge fluctuations instead of suppressing them, and local vertex corrections reshape screening in a state-dependent manner. A method that could only describe one regime would be a niche tool; one that reproduces both an insulator and a correlated metal without retuning is a framework.
Where this lands in the stack
The team positions the result as a “practical and reliable ab initio framework” for materials where nonlocal corrections beyond GW stay weak — the regime covering metal-insulator transitions and non-Fermi liquid behavior. That turns GW+EDMFT from a demonstration into a replacement candidate for the constrained-RPA step in existing DFT+DMFT pipelines, the same pipelines computational groups run nightly on national-lab supercomputers.
What is not yet shown, on the evidence in the report, is performance scaling beyond the two test materials or a published production implementation for third-party users — those remain open questions the sources do not address.
The Flatiron Institute, where much of this work was carried out, is the internal research division of the Simons Foundation launched in 2016 and named for the Flatiron District of New York City, where it is based. Its mission is to advance scientific research through computational methods, and it is organized into centers devoted to computational astrophysics, biology, quantum physics, mathematics and neuroscience.
The work described here falls under the Center for Computational Quantum Physics, launched in 2017, whose mission is to solve the quantum many-body problem and use the solutions to predict the behavior of materials and molecules of scientific and technological interest.
Yeh, Petocchi, Hampel, Werner, Parcollet, Georges and Morales published the work as “Electronic correlations and dynamical screening with ab initio quantum embedding.”