Development of Neural Network Force Fields with Innovative Reliability and Their Application to 2D Layered Materials

2023 Fiscal Year Final Research Report

• Project/Area Number: 21K17752
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 60100:Computational science-related
• Research Institution: National Institute for Materials Science (2022-2023); Japan Advanced Institute of Science and Technology (2021)
• Principal Investigator: NAKANO Kosuke 国立研究開発法人物質・材料研究機構, マテリアル基盤研究センター, 独立研究者 (50870903)
• Project Period (FY): 2021-04-01 – 2024-03-31
• Keywords: 第一原理量子モンテカルロ法 / 機械学習力場 / 第一原理電子状態計算

Outline of Final Research Achievements

In this study, we expanded the application of the ab initio Quantum Monte Carlo method, which precisely solves the many-body Schrodinger equation, to include 'multi-scale simulations' across multiple spatial scales. This research reveals that the accuracy pursued by the applicant at the electronic structure level plays a crucial role even at larger spatial scales. Throughout the research period, I established technologies for constructing a framework for machine learning force fields that enable molecular dynamics calculations with the accuracy of the ab initio Quantum Monte Carlo method and high-throughput computation techniques. Using the phase diagram of high-pressure hydrogen as a target system, I clarified the role that accuracy at the electronic structure level plays at larger spatial scales, achieving the initial objective of the study.

Free Research Field

電子状態計算

Academic Significance and Societal Importance of the Research Achievements

分子動力学計算の精度を決定づけるのは力場の精度であり, その構築には, 学習データが必要である. 固体周期系では, 密度汎関数法を超える高精度データ生成器としては第一原理量子モンテカルロ法が唯一の選択肢であったが, その計算データを利用した分子動力学計算用の力場の構築は報告されていなかった. 本研究成果は, その第一原理量子モンテカルロ法を利用して生成した信頼性の高い訓練データを利用して, 機械学習力場を構築する技術を確立したものである, 今後の, 計算材料科学に基づく物質の性質予測の定量性向上に貢献する.