Application of Neural Network Driven Molecular Dynamics with First-Principles Accuracy to Origin of Life

2021 Fiscal Year Final Research Report

• Project/Area Number: 19K14676
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 13040:Biophysics, chemical physics and soft matter physics-related
• Research Institution: Kumamoto University
• Principal Investigator: Shimamura Kohei 熊本大学, 大学院先端科学研究部(理), 助教 (60772647)
• Project Period (FY): 2019-04-01 – 2022-03-31
• Keywords: 機械学習力場（原子間ポテンシャル） / 人工ニューラルネットワーク / 分子動力学法 / 第一原理分子動力学法 / 生命の起源

Outline of Final Research Achievements

In this study, we refined a machine learning-based force field construction method to investigate chemical reactions in highly nonequilibrium systems such as deep-sea hydrothermal vent environments from a microscopic viewpoint. Although there is still room for improvement in describing chemical reactions in highly nonequilibrium systems, we have gained fruitful knowledge of useful construction methods. The atomic forces and pressure in addition to potential energy were trained to improve the accuracy of molecular dynamics simulations. The trained force fields accomplished not only describing solid-liquid phase transitions but also calculating thermal conductivity. It was also found that a more robust force field can be constructed by adjusting the coefficient of the cost function. We have suggested the worsening accuracy of the force field due to the arbitrariness of the atomic energy, and have proposed a possible solution to the issue with a data-driven approach.

Free Research Field

計算物性物理学

Academic Significance and Societal Importance of the Research Achievements

機械学習に基づく力場は低計算コストながら第一原理計算の精度を有するため、空間・時間スケールの観点でこれまで手が届かなかった領域へ我々を導く潜在力があり、物性分野においてフロンティアを開拓しつつある。それゆえ、潜在力をより物理的かつ実用的なレベルに押し上げるために、深海熱水噴出孔環境などの難易度の高い計算対象に取り組むことで、本研究では様々な重要な構築方法の要素を世に示せたと考えている。非平衡性の高い系に対するアプローチを完全に確立できたわけではないが、原子エネルギーという新たな視点から精度向上が期待できる手法を考案するなど、今後この分野に貢献する可能性のある成果も得られている。