Inverse materials design by integrating transfer learning techniques into a Bayesian framework

2021 Fiscal Year Final Research Report

• Project/Area Number: 18K18017
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 60030:Statistical science-related
• Research Institution: The Institute of Statistical Mathematics
• Principal Investigator: Wu Stephen 統計数理研究所, データ科学研究系, 准教授 (70804186)
• Project Period (FY): 2018-04-01 – 2022-03-31
• Keywords: polymer informatics / transfer learning

Outline of Final Research Achievements

Based on two key machine learning technologies, Bayesian inference and transfer learning, I have developed an all-in-one materials informatics platform in Python, called XenonPy, that can produce candidates of materials structures with desired properties. The openly available platform serves as the foundation to promote using data science for accelerating discovery of new functional materials suitable for various industrial needs in our daily life. Using this platform, three new high thermal conductivity polymers were discovered and successfully synthesized. The best performing design reached 80% higher thermal conductivities than conventional commercial polymers. Furthermore, several new liquid-crystal polyimide and high lattice thermal conductivity crystals were discovered that seem to have high industrial values.

Free Research Field

統計

Academic Significance and Societal Importance of the Research Achievements

材料科学と機械学習の融合を図るマテリアルズインフォマティクス呼ばれる学際領域に期待が高まっているが、産業的に重要かつ新規な機能性材料の合成に成功した研究はまだ少ない。実験コストが高いためにデータが不足していることが、マテリアルインフォマティクスの実際の産業価値を実現するための重要なボトルネックになっている。同技術を用いることで，データが豊富な重要度の低い材料特性から、データが少ない重要度の高い材料特性にも関連する有用な情報を抽出することができる。今までデータ不足のためマテリアルインフォマティクスの応用が失敗した材料科学問題を突破する鍵になる研究である。