ホーム > 研究活動 > 口頭発表(2019) > A Machine-Learning Scheme for Searching New Rare-Earth Magnet Compounds

研究活動

A Machine-Learning Scheme for Searching New Rare-Earth Magnet Compounds

Sixth Japan-U.S. Bilateral Meeting on Rare Metals, ワシントンDC (アメリカ)

2019年1月16日(水)

Taro Fukazawa (AIST)

Abstract

 Application of machine learning to a first-principles dataset has attracted much attention these days. Researchers have already made some attempts to find new magnet compounds with using a big dataset. We also are developing a scheme based on Bayesian optimization. We take the example of optimization of magnetization and the Curie temperature of (R1-αZ α)(Fe1-βCoβ)12-γTiγ that has the ThMn12 structure (R = Y, Nd, Sm; Z = Zr, Dy). In this search, we find that picking 100 systems with the Bayesian scheme out of 3630 systems on a candidate list is enough to get the top 10 systems of a magnetic property. Its efficiency can severely depend on a choice of the descriptor, a form to which systems considered are decoded. We show that the efficiency of a search is largely improved when the descriptor is appropriately chosen while bad descriptors lead to worse results than that with random choice.


研究活動

文部科学省

文部科学省
元素戦略プロジェクト(活動紹介)
NIMS磁石パートナーシップ

元素戦略拠点

触媒・電池元素戦略拠点
触媒・電池元素戦略研究拠点 (京都大学)
東工大元素戦略拠点
東工大元素戦略拠点 (東京工業大学)
構造材料元素戦略研究拠点
構造材料元素戦略研究拠点 (京都大学)
高効率モーター用磁性材料技術研究組合
高効率モーター用 磁性材料技術研究組合