Transfer learning to predict the properties of compounds: pre-trained model library

2022 Fiscal Year Final Research Report

• Project/Area Number: 20K19866
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 61030:Intelligent informatics-related
• Research Institution: The Institute of Statistical Mathematics
• Principal Investigator: Liu Chang 統計数理研究所, ものづくりデータ科学研究センター, 特任助教 (30814149)
• Project Period (FY): 2020-04-01 – 2023-03-31
• Keywords: 転移学習 / スモールデータ / 物性予測 / 結晶構造予測

Outline of Final Research Achievements

In this study, we have successfully employed transfer learning, a powerful machine learning technique, to address the challenge of limited material data and enable predictive extrapolation in the field of materials informatics (MI).

Our research has yielded significant outcomes. Firstly, we have developed XenonPy.MDL, an extensive model library containing a multitude of trained models. This library serves as a valuable resource for further advancements in the field. Secondly, we have applied transfer learning on thermodynamic stability prediction of high-entropy alloys and lattice thermal conductivity prediction, leading to notable findings presented at prestigious international conferences. Furthermore, our work has extended into crystal structure prediction, where we have introduced prediction algorithms surpassing the performance of conventional methods. We have achieved remarkable results by proposing an innovative prediction algorithm, particularly in crystal structure prediction.

Free Research Field

マテリアルズインフォマティクス

Academic Significance and Societal Importance of the Research Achievements

学術的意義としては，転移学習の技術はデータ収集に高いコストがかかる材料研究分野において必要不可欠である．本研究で挙げられたハイエントロピー合金の熱力学安定性予測と結晶構造予測の研究成果はその実例であり，小規模な第一原理計算の結果のみで高精度な予測を実現した．転移学習技術の導入は．研究の効率化と新たな技術の実現に向けた重要な一歩となった．
社会的意義としては，転移学習の導入により，材料の設計や特性予測の精度と外挿性能が向上し，材料開発のスピードが加速されることが期待できる．これにより，エネルギー効率の高い材料や環境負荷の低い製品の開発が促進され，持続可能な社会の実現に寄与することができる．