プレプリント / バージョン1

AI-agent-driven development of scientific software: a case study with determinant quantum Monte Carlo

##article.authors##

DOI:

https://doi.org/10.51094/jxiv.6717

キーワード:

generative AI、 AI agent、 scientific software、 determinant quantum Monte Carlo、 strongly correlated electron systems

抄録

AI coding agents, which autonomously plan, generate, execute, and debug code under human supervision, are rapidly changing software development, including in the scientific community. Here, we examine the capability and reliability of AI coding agents in developing research-grade scientific software from scratch during the three-day hackathon AIMHack2026. As the target, we chose to implement the determinant quantum Monte Carlo (DQMC) method for the half-filled Hubbard model on bipartite lattices. For this system, DQMC yields numerically exact finite-temperature results, apart from statistical errors, in the zero-time-step limit. For this target, exact diagonalization, the U = 0 analytic solution, and Trotter extrapolation provide a series of quantitative validation tests. Within three days, the agents produced a C implementation with numerical stabilization and parallelization over independent Monte Carlo chains. We confirm that the results obtained with this code are consistent with exact-diagonalization results. The development process was recorded in logs with date and model attribution, allowing us to trace how the agents built and validated the code. The source code and development logs are publicly available in the SAI-QMC repository on GitHub. We find that two preparations were important for this outcome, namely (i) a detailed algorithmic reference prepared in advance and (ii) test-driven development organized around this series of validation tests. Since agentic coding itself will continue to evolve rapidly, this paper provides a validated, quantitative record of what AI coding agents achieved in scientific software development at this time.

利益相反に関する開示

No potential conflict of interest was reported by the author.

ダウンロード *前日までの集計結果を表示します

ダウンロード実績データは、公開の翌日以降に作成されます。

引用文献

OpenAI. ChatGPT ; 2026. Accessed: 2026-08-08; Available from: https://chatgpt.com.

GitHub. GitHub Copilot ; 2026. Accessed: 2026-09-28; Available from: https://github.com/features/copilot.

Anysphere. Cursor ; 2026. Accessed: 2026-09-28; Available from: https://cursor.com/.

Anthropic. Claude Code ; 2026. Accessed: 2026-08-08; Available from: https://claude.com/claude-code.

OpenAI. Codex ; 2026. Accessed: 2026-08-08; Available from: https://openai.com/codex/.

Yoshimi K, Terasaki S, Yamada G. Building a research-software catalog with a coding agent: from hackathon prototype to public deployment [arXiv:2609.04711]; 2026. Available from: https://arxiv.org/abs/2609.04711.

Misawa T, Koizumi A, Tamura R, et al. Exploring utilization of generative AI for research and education in data-driven materials science. Science and Technology of Advanced Materials: Methods. 2025;5(1):2535956.

Blankenbecler R, Scalapino DJ, Sugar RL. Monte Carlo calculations of coupled boson-fermion systems. I. Phys Rev D. 1981;24:2278–2286.

Hirsch JE. Discrete Hubbard-Stratonovich transformation for fermion lattice models. Phys Rev B. 1983;28:4059–4061.

Anthropic. Models overview — Claude Docs ; 2026. Accessed: 2026-08-12; Available from: https://docs.claude.com/en/docs/about-claude/models.

OpenAI. Models — OpenAI API Documentation ; 2026. Accessed: 2026-08-12; Available from: https://platform.openai.com/docs/models.

Google. Gemini Models — Google AI for Developers ; 2026. Accessed: 2026-08-12; Available from: https://ai.google.dev/gemini-api/docs/models.

Liu JG. Vibe coding done right [Lecture at CompPhysHack2026, Naha, Japan, March 2026]; 2026. Accessed: 2026-08-08; Available from: https://github.com/qc-hybrid/CompPhysHack2026.

Vincent J. Superpowers: Agentic Skills Framework for Claude Code ; 2025. Accessed: 2026-08-12; Available from: https://github.com/obra/superpowers.

Tasaki H. Physics and Mathematics of Quantum Many-Body Systems. Cham: Springer; 2020. Available from: https://doi.org/10.1007/978-3-030-41265-4.

Hirsch JE. Two-dimensional Hubbard model: Numerical simulation study. Phys Rev B. 1985;31:4403–4419.

White SR, Scalapino DJ, Sugar RL, et al. Numerical study of the two-dimensional Hubbard model. Phys Rev B. 1989;40:506–516.

Otsuka Y, Yunoki S, Sorella S. Universal quantum criticality in the metal-insulator transition of two-dimensional interacting Dirac electrons. Phys Rev X. 2016;6:011029.

Kawamura M, Yoshimi K, Misawa T, et al. Quantum lattice model solver HΦ. Comput Phys Commun. 2017;217:180–192.

Otsuka Y. Effects of disorder and interactions in electron systems on lattices [in Japanese] [dissertation]. Department of Applied Physics, University of Tokyo; 2002. Appendix A available online. Accessed: 2026-10-01; Available from: https://github.com/otsukay/thesis.

Finite-temperature auxiliary-field quantum Monte Carlo (DQMC/BSS): C implementation design [in Japanese] [SAI-QMC development documentation]; 2026. Accessed: 2026-09-23; Available from: https://github.com/tmisawa/SAI-QMC/blob/main/docs/superpowers/specs/2026-06-24-finite-T-aux-field-qmc-design.md.

Finite-temperature auxiliary-field quantum Monte Carlo (DQMC/BSS): implementation plan [in Japanese] [SAI-QMC development documentation]; 2026. Accessed: 2026-09-23; Available from: https://github.com/tmisawa/SAI-QMC/blob/main/docs/superpowers/plans/2026-06-24-finite-T-aux-field-qmc.md.

Misawa T. SAI-QMC [GitHub repository]; 2026. Accessed: 2026-09-23; Available from: https://github.com/tmisawa/SAI-QMC.

ダウンロード

公開済


投稿日時: 2026-10-05 04:50:52 UTC

公開日時: 2026-11-16 06:52:06 UTC
研究分野
物理学