Neural networks as low-cost surrogates for impurity solvers in quantum embedding methods

arXiv:2603.25557 · cond-mat.str-el · Submitted 2026-03-26 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Quantum Radio. Generated commentary on the latest quantum physics and condensed matter papers.

Kai: Today's paper: "Neural networks as low-cost surrogates for impurity solvers in quantum embedding methods".

Mira: A promising application of machine learning is the creation of low-cost surrogate models to mitigate computational bottlenecks in quantum many-body simulations.

Kai: First, who's behind it and why it matters.

Title and authors: Kai: So, we're looking at the paper titled "Neural networks as low-cost surrogates for impurity solvers in quantum embedding methods," and it’s written by Rohan Nain, Philip M. Dee, Kipton Barros, Steven Johnston, and Thomas A. Maier. What does that title really suggest about what they’ve been doing?

Mira: It suggests they are using neural networks to create these lower-cost models to help with the bottlenecks in quantum many-body simulations which is a big deal for complex problems like correlated electron models.

Lev: From my side, if this works on a simulator, the next thing we need is knowing how robust it is when we try to run it on actual quantum hardware; that’s where I get my concerns.

Kai: Exactly, and looking at the authors' backgrounds, Nain and Dee are from physics and astronomy departments at the University of Tennessee Knoxville, while Johnston and Maier are with Oak Ridge National Laboratory in computational sciences; that mix tells us this is a joint effort between theory and serious computational engineering.

Mira: It sounds like they are tackling a problem where we usually have to use very expensive numerical methods for the impurity solver within dynamical mean field theory simulations of these correlated electron models.

Lev: That’s exactly where the cost comes in; if this AI can replace something that scales poorly with system size or temperature, that opens up a lot more possibilities for what we can actually simulate on real hardware.

The paper's summary: Kai: So, the core of this paper describes how they use a neural network to act as an efficient substitute for the impurity solver in dynamical mean field theory simulations, aiming to reduce computational bottlenecks significantly.

Mira: The summary explains that their neural network solver manages to achieve accuracy levels comparable to popular continuous-time quantum Monte Carlo solvers when it’s interpolating between the samples they trained on.

Lev: Achieving accuracy comparable to CT-QMC is impressive, but I need details on *how* that interpolation works; are we talking about a smooth transition or something more discrete?

Kai: The summary points out that this NN solver is designed to take the Weiss mean-field Green’s function, the inverse temperature, and the Hubbard interaction strength as inputs to predict the impurity site’s self-energy.

Mira: That prediction is trained using a physics-informed loss function, which means they aren't just guessing; they are training it with rules derived from quantum mechanics to minimize that error.

Lev: Minimizing error is one thing, but I wonder about the data efficiency mentioned in the summary; can this model actually be trained with fewer examples than previous high-fidelity solvers?

Kai: The summary highlights that they show this NN solver achieves accuracy comparable to popular continuous-time quantum Monte Carlo solvers within the interpolation region of their training set, and it converges in a comparable number of dynamical mean field theory iterations.

Mira: That convergence speed is really interesting because it suggests the simulation loop itself becomes much faster when using this surrogate model instead of running the full QMC solver every time.

Lev: Faster iteration counts are good for large systems, but what about when we try to push those simulations into regimes where they don't have good training data?

The paper's improvements: Kai: The paper lays out a few key improvements for this NN approach, focusing on how it handles extrapolation and generalization outside the original training distribution.

Mira: They show that the NN solver can extrapolate well for a range of temperatures below the training set coverage, meaning it can predict things even when those conditions weren't explicitly in their synthetic data.

Lev: Extrapolation is risky; if the physics changes abruptly outside your training bounds, how reliable is that prediction really? What happens if we need to push the temperature much further down than they tested?

Kai: They demonstrate that while model errors can become significant with extrapolation, imperfect predictions can still be useful as an initial guess to accelerate convergence toward self-consistency in traditional dynamical mean field theory solvers.

Mira: That capability suggests the NN isn't just a final answer generator; it acts as a powerful tool to significantly speed up the process of getting the full self-consistent solution from DMFT.

Lev: So, they’re proposing a hybrid workflow where we use this fast AI prediction first, and then run a more rigorous solver on top of that result when we step outside the known parameter space?

Kai: Precisely; they show that this hybrid approach can reduce wall time by factors of three point four times and five point seven times compared to pure continuous-time QMC convergence in those extrapolated regions.

Mira: That reduction in computation time is substantial, especially when you consider how many iterations a full QMC solver usually takes to reach convergence, which the paper implies is very costly at low temperatures and on large systems.

Lev: A factor of five reduction sounds like a lot when we think about running these simulations on actual quantum hardware where every second counts.

Conclusion: Kai: So, to wrap up the discussion on "Neural networks as low-cost surrogates for impurity solvers in quantum embedding methods," the main point is that a compact neural network can provide an accurate solution for a dynamical mean field theory simulation of the half-filled single-band Hubbard model.

Mira: They’ve shown this NN solver can accurately predict the impurity self-energy by training on a relatively small synthetic dataset, which is much more data efficient than prior work.

Lev: I'd add that their ability to produce high-fidelity predictions for observables like double occupancy and quasiparticle weight Z with mean absolute errors below zero point zero zero five confirms the predictive power of this approach even in those more detailed metrics.

Kai: And they’ve shown this model can map phase boundaries determined by the CT-QMC solver well in the interpolation region and qualitatively in the extrapolation regime just outside of it, which is really important for exploring parameter space quickly.

Mira: The implication here is that we might be able to rapidly explore large regions of parameter space where traditional methods are computationally prohibitive, using this surrogate model as a fast initial guess for more expensive exact solvers.

Lev: If this approach holds up when we start mapping these models onto real hardware, it means we could significantly reduce the time required for error characterization and system tuning in experiments.

Kai: Exactly; the paper demonstrates that using AI to tackle computational bottlenecks in quantum many-body simulations is a viable path forward, especially for systems like the Hubbard model where exact solvers are traditionally too slow.

Mira: This work shows that we don't necessarily need massive datasets to get good results if we use intelligent methods like this NN solver as a surrogate for complex impurity calculations.

Lev: I think the real impact is showing a concrete pathway to making these complex quantum simulations more accessible by drastically lowering the computational barrier, which is what matters most for experimentalists and error correction researchers.

Department of Physics and Astronomy, The University of Tennessee, Knoxville · Computational Sciences and Engineering Division, Oak Ridge National Laboratory · Theoretical Division and CNLS, Los Alamos National Laboratory

cond-mat.str-el

Submitted: 2026-03-26

Updated: 2026-04-23

Comments: 10 pages, 9 figures, v2 includes additional results for a larger training set

DOI: 10.1103/9ndg-shty

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 77/100

The gist: A promising application of machine learning is the creation of low-cost surrogate models to mitigate computational bottlenecks in quantum many-body simulations.

Key concepts

Dynamical Mean Field Theory (DMFT): DMFT is a powerful method used to study strongly correlated electron systems, such as the Mott transition. It simplifies complex many-body problems by mapping them onto an effective single-site problem embedded in a self-consistent bath. The accuracy of DMFT heavily relies on accurately solving the impurity problem within this framework.
Impurity Solver: The impurity solver is the computationally demanding component of DMFT that calculates the interacting Green’s function for a quantum impurity. Traditional methods like continuous-time QMC are exact but very slow, especially at low temperatures or for complex multi-orbital systems, making it a major computational bottleneck.
Neural Network Surrogate Model: This is an artificial intelligence model (a neural network) trained to mimic the behavior of a high-accuracy physical solver. It takes inputs like temperature and interaction strength and predicts the self-energy, effectively acting as a fast, low-cost substitute for the slow, traditional impurity solver.

Terminology

Regularization — 16× in this paper · explained in 3 episodes
Accuracy — 14× in this paper · explained in 1 episode
Hubbard Model — 13× in this paper · explained in 3 episodes
Consistency — 10× in this paper · explained in 3 episodes
Self-consistency — 10× in this paper · explained in 2 episodes
Generalization — 4× in this paper · explained in 13 episodes
Fidelity — 3× in this paper · explained in 2 episodes
Hybrid Approach — 3× in this paper · explained in 4 episodes
Hysteresis — 3× in this paper · explained in 3 episodes
Loss Function — 3× in this paper · explained in 2 episodes
Core — 2× in this paper · explained in 2 episodes
Dynamical Mean Field Theory (DMFT) — 2× in this paper · explained in 2 episodes
Ab Initio — 1× in this paper · explained in 1 episode
Fine-Tuning — 1× in this paper · explained in 5 episodes
Iterative Refinement — 1× in this paper · explained in 3 episodes
Scaffolding — 1× in this paper · explained in 3 episodes
Toolbox for Research on Interacting Quantum Systems — 1× in this paper
Transferability — 1× in this paper · explained in 4 episodes
Transformer — 1× in this paper · explained in 2 episodes
Transformer Architecture — 1× in this paper · explained in 2 episodes
US Department of Energy — 1× in this paper

Summary

A promising application of machine learning is the creation of low-cost surrogate models to mitigate computational bottlenecks in quantum many-body simulations. This research explores whether a neural network (NN) can be trained in the low-data regime, with one to two orders of magnitude fewer training examples than previous works, as an efficient substitute for the impurity solver in dynamical mean-field theory simulations of correlated electron models.

The gist

A neural network solver achieves accuracy comparable to popular continuous-time quantum Monte Carlo (CT-QMC) impurity solvers when interpolating between samples within the training set.

Background and Motivation

Quantum embedding methods, such as dynamical mean field theory (DMFT), are instrumental in shaping our understanding of strongly correlated phenomena like the Mott transition. The computationally expensive part of these simulations is the impurity solver, which calculates the interacting Green’s function G(τ) for a quantum impurity problem defined by the bath parameters. Traditionally, numerically exact methods like continuous-time QMC (CT-QMC), exact diagonalization (ED), or numerical renormalization group (NRG) are used, but these can be computationally demanding, particularly at low temperatures and when generalized to multi-orbital problems. For instance, popular CT-QMC solvers scale as the cube of the inverse temperature and number of orbitals, making them costly at low temperatures and on large systems.

The Neural Network Architecture

The proposed neural network solver is designed to replace the impurity solver within the DMFT loop. The architecture consists of four fully-connected dense layers of neurons with GELU activations. This network accepts the Weiss mean-field Green’s function G0(τ), the inverse temperature β, and the Hubbard interaction strength U as input features, and it predicts the impurity site’s self-energy Σ(τ). The network is trained to minimize a physics-informed loss function:

**L = **

Data Generation for Training

The training data generation process involves generating synthetic data using the CT-QMC impurity solver for a fixed (U, β) grid. The features and targets are constructed as follows:

The input features (Xˆ32) are defined as: Ui, εi, Gω=0,2,…,58.
The target output (Yˆ30) is defined as: f ! 0, f ! 2, …, f ! 58, where the coefficients are the Legendre polynomial expansion coefficients of the self-energy Σ(τ).

This process generates a dataset of Ntrain = 500 training examples. The paper notes that this approach uses relatively small training datasets, which is a deliberate choice to test data efficiency and mitigate the expense of generating high-quality synthetic data for larger models.

Performance and Results

The NN solver demonstrates excellent agreement with the more costly CTHYBQMC solver in both metallic and insulating regimes within the interpolation region of the training set, with an RMSE of 10−3 at convergence. The NN achieves convergence in a comparable number of DMFT iterations, though it is approximately four orders of magnitude faster than the QMC solver. Furthermore, the model extrapolates well for a range of temperatures below the training set coverage; while model errors can become significant with extrapolation, imperfect predictions can be useful as an initial guess to accelerate convergence to self-consistency within traditional DMFT solvers.

Extrapolation and Generalization

The study examines performance in regions where the NN must extrapolate to lower temperatures outside the training distribution. The NN successfully reproduces phase boundaries determined by the CT-QMC solver well in the interpolation regime and qualitatively in the extrapolation regime just outside of it. The model's ability to generalize under moderate extrapolation suggests that its output can be efficiently refined using an exact solver, such as running a new DMFT loop seeded by the NN output and switching to CTHYB QMC. This hybrid approach significantly reduces wall time, showing a reduction in computation time by factors of 3.4x and 5.7x compared to pure CT-QMC convergence in the extrapolated region.

Dataset Size Scaling

To assess sensitivity, a second NN surrogate was trained with Ntrain = 1600 samples. This larger dataset resulted in significantly better quantitative accuracy for observables like double occupancy D and quasiparticle weight Z, reducing the mean absolute error (MAE) to below 0.005 in the metallic sweep at βt = 25, which is approaching CTHYB accuracy at a training cost that remains two orders of magnitude smaller than prior studies employing 16,000 samples. The results confirm that increasing the training set size improves predictive power across both interpolation and extrapolation regimes.

Conclusion

A compact NN-based surrogate trained on a relatively small synthetic dataset can provide an accurate solution for a DMFT simulation of the half-filled single-band Hubbard model by predicting the impurity self-energy Σ(τ).

Improvements for AI systems

Here are specific improvements for AI systems based on the findings of this research:

The ability to train a Neural Network (NN) surrogate solver to replace computationally expensive impurity solvers (like CT-QMC) in Dynamical Mean-Field Theory (DMFT) simulations of correlated electron models, specifically the half-filled Hubbard model.
The NN solver's capability to achieve accuracy comparable to high-fidelity QMC solvers when interpolating between training samples, allowing for a significant reduction in computational time (up to a factor of five).
The ability for the NN solver to serve as an efficient warm start or initial guess for more accurate, traditional impurity solvers when extrapolating into regions outside the training distribution, thereby accelerating convergence to self-consistency.
The development of a data-efficient ML framework where a relatively small synthetic training dataset (e.g., 500 samples) is sufficient to train an NN surrogate to predict the impurity self-energy, rather than predicting the Green's function directly.
The ability of the trained NN model to accurately map and reproduce complex phase diagrams (e.g., Mott metal-insulator transitions) across both interpolation and moderate extrapolation regimes in parameter space, allowing for rapid exploration of large regions that would be computationally prohibitive with traditional methods.
The capacity for the NN system to produce high-fidelity predictions of key physical observables, such as double occupancy and quasiparticle weight (Z), with mean absolute errors significantly lower than those obtained from prior ML studies using larger datasets (e.g., achieving MAEs below 0.005 in double occupancy).
The capability for the NN system to perform hybrid acceleration, where its fast prediction is used to seed a subsequent, more expensive exact solver (CT-QMC), leading to substantial wall-time reductions (up to a factor of 3.4 reduction in wall time for convergence in the insulating regime).

Abstract

A promising application of machine learning is the creation of low-cost surrogate models to mitigate computational bottlenecks in quantum many-body simulations. Here, we explore whether a neural network (NN) can be trained in the low-data regime, with one to two orders of magnitude fewer training examples than previous works, as an efficient substitute for the impurity solver in dynamical mean-field theory simulations of correlated electron models. We show that the NN solver achieves accuracy comparable to popular continuous-time quantum Monte Carlo (CT-QMC) impurity solvers when interpolating between samples within the training set. While the NN's performance decreases notably when extrapolating to lower temperatures outside the training distribution, its output still provides an excellent initial guess for input to more accurate CT-QMC impurity solvers, thus accelerating the time to solution up to a factor of five. We discuss our results in the context of rapid phase-space exploration.