Enabling large language model agents for high-fidelity many-body perturbation theory calculations
Abstract
Many-body perturbation theory (MBPT) methods, including the GW approximation, the GW plus Bethe-Salpeter equation (GW-BSE) method, and GW perturbation theory (GWPT), among others, provide accurate first-principles descriptions of excited-state properties and multiparticle interactions of materials. However, their use remains limited due to complex multi-software workflows, interdependent physical parameters, and substantial theoretical expertise. Large language model (LLM) agents, which couple a language model to software tools, show potential in automating computational tasks, but MBPT remains an unaddressed challenge. We show that LLM agents tested in our work, including frontier models such as Opus 4.7 and GPT-5.5, do not reliably construct MBPT workflows in one attempt, suggesting that correct setup requires expert knowledge and consistency that current models lack. Across tested model-harness combinations, out-of-the-box agents average 54%, 36%, and 19% on GW, GW-BSE, and GWPT workflow benchmarks, with strong model dependence and large run-to-run variation. Errors typically originate from cross-stage inconsistencies that surface only several stages after they are introduced. We therefore develop AI-Donkey, an agentic framework that combines expert-curated knowledge with deterministic execution to enable high-fidelity GW, GW-BSE, and GWPT calculations and analyses. With AI-Donkey, all tested agents, from frontier proprietary models to open-weight models, produce complete, runnable MBPT workflows with no observed run-to-run variation. We further demonstrate autonomous calculations and post-processing on 1D, 2D, and bulk materials, including the many-body enhancement of Fröhlich electron-phonon coupling in polar semiconductors. Our approach provides a general framework for knowledge engineering to enable reliable agentic workflows in computational materials science.
Related
- DefectSkills: Guiding AI agents through first-principles calculations for point defects in semiconductors and insulators
- Can Autonomous LLM Agents Execute Multireference Quantum Chemistry Calculations?
- MedPhysBench: A Benchmark for Evaluating Language Models on Medical Physics Calculations and Error Detection
- Toward a social psychology of AI: language-model agents reproduce human-like minimal-group bias
- Quantum Encoding Agents: A Natural Language Interface for Data Embedding Strategy Selection in Quantum Machine Learning
BibTeX
@misc{zhang2026enabling,
title = {{Enabling large language model agents for high-fidelity many-body perturbation theory calculations}},
author = {Benran Zhang and Byeongchan Lee and Zien Zhu and Chih-En Hsu and Zhenfa Zheng and Hung-Chung Hsueh and Young Woo Choi and Mauro Del Ben and Zhenglu Li},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.20779},
eprint = {2610.20779},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.20779},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Enabling large language model agents for high-fidelity many-body perturbation theory calculations
AU - Benran Zhang
AU - Byeongchan Lee
AU - Zien Zhu
AU - Chih-En Hsu
AU - Zhenfa Zheng
AU - Hung-Chung Hsueh
AU - Young Woo Choi
AU - Mauro Del Ben
AU - Zhenglu Li
PY - 2026
DA - 2026/10/14
PB - arXiv
DO - 10.48550/arxiv.2610.20779
UR - https://arxiv.org/abs/2610.20779
AB - Many-body perturbation theory (MBPT) methods, including the GW approximation, the GW plus Bethe-Salpeter equation (GW-BSE) method, and GW perturbation theory (GWPT), among others, provide accurate first-principles descriptions of excited-state properties and multiparticle interactions of materials. However, their use remains limited due to complex multi-software workflows, interdependent physical parameters, and substantial theoretical expertise. Large language model (LLM) agents, which couple a language model to software tools, show potential in automating computational tasks, but MBPT remains an unaddressed challenge. We show that LLM agents tested in our work, including frontier models such as Opus 4.7 and GPT-5.5, do not reliably construct MBPT workflows in one attempt, suggesting that correct setup requires expert knowledge and consistency that current models lack. Across tested model-harness combinations, out-of-the-box agents average 54%, 36%, and 19% on GW, GW-BSE, and GWPT workflow benchmarks, with strong model dependence and large run-to-run variation. Errors typically originate from cross-stage inconsistencies that surface only several stages after they are introduced. We therefore develop AI-Donkey, an agentic framework that combines expert-curated knowledge with deterministic execution to enable high-fidelity GW, GW-BSE, and GWPT calculations and analyses. With AI-Donkey, all tested agents, from frontier proprietary models to open-weight models, produce complete, runnable MBPT workflows with no observed run-to-run variation. We further demonstrate autonomous calculations and post-processing on 1D, 2D, and bulk materials, including the many-body enhancement of Fröhlich electron-phonon coupling in polar semiconductors. Our approach provides a general framework for knowledge engineering to enable reliable agentic workflows in computational materials science.
ER -