🔬 mesoscale physics

Deep Learning GW Quasiparticle Hamiltonians for Many-Body Excited-State Electronic Structure at Scale

The paper introduces DeepH-GW, a deep-learning framework that predicts accurate, linear-scaling GW quasiparticle Hamiltonians directly from atomic structures, enabling large-scale many-body excited-state simulations with few-meV precision and strong cross-scale transferability.

Original authors: Xiaoxun Gong, Zechen Tang, Woochang Kim, Yang Li, Wenhui Duan, Yong Xu, Steven G. Louie

Published 2026-09-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xiaoxun Gong, Zechen Tang, Woochang Kim, Yang Li, Wenhui Duan, Yong Xu, Steven G. Louie

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

To understand how materials conduct electricity, emit light, or power the next generation of quantum computers, scientists must look beyond the quiet, stable state of atoms at rest. They must examine the excited states—the fleeting moments when electrons are jolted by energy and begin to move. For decades, the standard tool for predicting these behaviors has been a method called density functional theory. While efficient, this approach often fails to provide a quantitatively reliable picture of these excited electrons, treating them more like a cloud of probability than distinct particles with specific energies. To fix this, researchers use a more rigorous, high-fidelity approach known as the GW approximation. This method solves a complex equation that accounts for how electrons interact with one another, offering a precise map of their behavior. However, this precision comes at a steep price: the calculations are so computationally demanding that they can only be performed on small, simple structures. When scientists try to apply this method to large, complex materials or systems with thousands of atoms, the computational cost becomes prohibitive, leaving a gap between the accuracy needed for discovery and the power available to achieve it.

A team of researchers has now bridged this gap with a new framework called DeepH-GW. By combining the precision of high-level physics with the speed of modern artificial intelligence, they have created a system that can predict the behavior of excited electrons in massive materials with remarkable speed and accuracy. The core of their achievement is a deep-learning model that learns to predict an "effective Hamiltonian." In the language of physics, a Hamiltonian is a mathematical object that contains all the information needed to describe the energy and behavior of a system. The researchers trained their neural network on high-quality data generated from small, perfect crystals. Once trained, the model learned the rules governing how electrons interact within these structures. The surprising discovery was that these rules are local; the behavior of an electron in one part of a crystal is determined primarily by its immediate neighbors, not by the entire material. This "nearsightedness" allowed the model, trained on tiny structures, to be applied to much larger ones without losing accuracy.

The researchers tested this approach on two common semiconductors: diamond and gallium phosphide. They first verified that their method could reconstruct the complex electronic structure of these materials from scratch, matching the results of the most rigorous, traditional calculations. The neural network, which takes only the positions of atoms as input, successfully predicted the energy levels of electrons with an error of just a few thousandths of an electron volt. This level of precision is critical because it is small enough to capture subtle physical effects that often determine whether a material is useful for a specific technology. More importantly, the team demonstrated that the model could transfer its knowledge from small training sets to vastly larger supercells containing thousands of atoms. While traditional methods would require weeks or months of supercomputer time to analyze such large systems, the new framework performed these predictions in a matter of hours, scaling linearly with the size of the material.

One of the most compelling applications of this speed is the study of how temperature affects materials. In semiconductors, the gap between energy levels where electrons can exist changes as the material heats up, a phenomenon driven by the vibration of atoms, known as electron-phonon coupling. To calculate this accurately using traditional methods, scientists would need to simulate thousands of different atomic vibrations, a task that is currently impossible for large systems. Using DeepH-GW, the researchers simulated these thermal vibrations in a diamond crystal containing thousands of atoms. They found that their model could accurately predict how the band gap of diamond shrinks as the temperature rises, matching experimental data far better than previous, less accurate methods. This capability suggests that the framework can now be used to explore complex, real-world materials that were previously out of reach, such as twisted layers of graphene, quantum defects in diamonds, or interfaces between different materials.

The work does not claim to have solved every problem in electronic structure theory. The researchers note that their current model relies on specific approximations to keep the calculations fast and that it may require adjustments for materials with strong long-range electrical interactions, such as certain metals or polar insulators. They also point out that the accuracy of the final prediction depends on the quality of the initial training data and the specific mathematical basis used to represent the atoms. However, the results establish a clear path forward. By proving that a machine learning model can learn the complex, many-body rules of quantum mechanics from small examples and apply them to large, disordered systems, the team has provided a practical engine for large-scale simulation. This approach transforms the study of excited states from a bottleneck into a scalable process, opening the door to the discovery of new materials with engineered optical and quantum properties.

Technical Summary: DeepH-GW for Many-Body Excited-State Electronic Structure

Problem Statement
Accurate prediction of quasiparticle electronic structures is essential for understanding excited-state properties, optoelectronics, and transport phenomena in materials. While the $GW$ approximation within many-body perturbation theory provides a standard, high-fidelity framework for these calculations, it remains computationally prohibitive for large systems, complex atomic configurations, and high-throughput discovery. This creates a fundamental accuracy-efficiency dilemma: the systems requiring precise excited-state predictions (e.g., large supercells, disordered materials) are often inaccessible to conventional $GW$ methods due to their steep computational scaling. Existing machine learning approaches have accelerated ground-state DFT calculations but have struggled to extend to the quasiparticle level, primarily due to the intrinsic nonlocality of the $GW$ self-energy and the difficulty of generating high-fidelity training data.

Methodology: DeepH-GW Framework
The authors introduce DeepH-GW, a deep-learning framework designed to predict an effective $GW$ quasiparticle Hamiltonian directly from atomic structure. The methodology bridges the gap between high-fidelity plane-wave (PW) $GW$ calculations and scalable, localized neural network inference through three core components:

Real-Space Hamiltonian Reconstruction: To reconcile the nonlocal nature of $GW$ interactions with the locality required for scalable message-passing neural networks, the framework constructs an effective static quasiparticle Hamiltonian ( $H^{GW}$ ) expressed in a localized atomic orbital (AO) basis. This is achieved by projecting converged PW $GW$ results (specifically one-shot $G_0W_0$ ) onto an AO basis using a reconstruction interface. The effective Hamiltonian incorporates a static, Hermitian self-energy ( $\Sigma^{stat}$ ), derived from the frequency-dependent self-energy, which allows for the diagonalization of the Hamiltonian to yield quasiparticle energies.
E(3)-Equivariant Architecture: The framework utilizes the DeepH neural network architecture, which employs equivariant message-passing layers. This design respects the fundamental spatial symmetries (Euclidean group E(3)) and exploits the "nearsightedness" of electronic matter. The model learns the mapping from local atomic environments to the matrix elements of the effective $GW$ Hamiltonian.
Training and Inference Strategy: The models are trained on high-fidelity PW $G_0W_0$ data for relatively small supercells (e.g., $3 \times 3 \times 3$ ). Crucially, the approach does not rely on lower-level electronic descriptors (like DFT wavefunctions) as inputs to preserve linear-scaling inference; it uses only atomic structure. The reconstruction process allows the use of systematically convergent PW data while enabling the final prediction to be performed via efficient, linear-scaling sparse diagonalization of the predicted AO Hamiltonian.

Key Contributions and Results
The paper demonstrates the efficacy of DeepH-GW through several benchmarks on diamond and Gallium Phosphide (GaP):

Hamiltonian Reconstruction Accuracy: The authors validated that the reconstructed AO effective Hamiltonians faithfully reproduce the quasiparticle band structures obtained from original PW $G_0W_0$ calculations. While minor deviations exist (attributed to finite AO basis sets and the truncation of the corrected band manifold), the reconstruction is sufficiently accurate to serve as a training target.
Cross-Scale Transferability: A primary achievement is the demonstration of transferability from small training structures to substantially larger supercells. Models trained exclusively on $3 \times 3 \times 3$ $3 \times 3 \times 3$ supercells were successfully applied to $6 \times 6 \times 6$ $6 \times 6 \times 6$ supercells.
- For diamond, the mean absolute error (MAE) for the 20 bands closest to the Fermi level was 5.9 meV.
- For GaP, the MAE for the 40 closest bands was 9.7 meV.
- These results confirm that the effective $GW$ Hamiltonian retains sufficient nearsightedness to support deep learning across scales, despite the long-range nature of many-body interactions.
Electron-Phonon Renormalization: The framework was applied to calculate the temperature-dependent band-gap renormalization of diamond due to electron-phonon coupling. By generating thermal displacement configurations in large ( $12 \times 12 \times 12$ ) supercells and predicting their quasiparticle Hamiltonians, the authors computed spectral functions and band gaps. The DeepH-GW results showed substantially better agreement with experimental data than standard DFT-level calculations, capturing many-body corrections that are typically too expensive to compute directly for such large systems.

Significance and Claims
The paper claims that DeepH-GW provides a practical route toward large-scale many-body simulations and the development of foundation models for excited-state electronic structure. By combining the systematic accuracy of plane-wave methods with the linear-scaling efficiency of neural network inference, the framework overcomes the historical bottleneck preventing $GW$ calculations on systems with thousands of atoms.

The authors emphasize that this approach enables the study of phenomena in moiré supercells, quantum defects, junctions, and heterogeneous materials that were previously inaccessible. Furthermore, the successful prediction of electron-phonon band-gap renormalization illustrates the framework's capability to capture complex many-body physics in thermally displaced systems. The work establishes a pathway for autonomous discovery workflows for materials with engineered excitonic and optoelectronic functionalities, while acknowledging that future improvements in basis set optimization and the inclusion of diverse material classes (metals, low-dimensional systems) are necessary for broader applicability.

Technical Summary: DeepH-GW for Many-Body Excited-State Electronic Structure

More like this