2021 Fiscal Year Research-status Report
Query-and-Learn Machine Learning framework to model the stability mechanism of REFe12 magnets
| Project/Area Number |
21K14396
|
|---|
| Research Institution | Japan Advanced Institute of Science and Technology |
|---|
Principal Investigator |
NGUYEN DuongNguyen 北陸先端科学技術大学院大学, 先端科学技術研究科, 助教 (20879978)
|
|---|
| Project Period (FY) |
2021-04-01 – 2024-03-31
|
| Keywords | rare-earths magnets / active learning / SmFe12 / materials discovery |
|---|
| Outline of Annual Research Achievements |
We propose a query-and-learn active learning combined with first-principles calculations to search for stable SmFe12 structures with ThMn12 skeleton via substitution method and clarify their stabilization mechanism.
3307 SmFe120-α-βXαYβ structures are prepared by substituting X,Y elements Mo,Zn,Co,Cu,Ti,Al,Ga with α+β<4 into Fe sites of the original SmFe12 structure.
Our machine learning model get prediction error of formation energy at 1.25×10-2 eV/atom using only 1/6 training data compared to other methods. The optimal recall rate for stable structures is 4-times faster than the random search. The formation energy landscape visualized using the embedding representation revealed that the substitutions of Al and Ga have the highest potential to stabilize the SmFe12 structure. In particular, SmFe9[Al/Ga]2Ti showed the highest stability amongst the investigated structures. Also, the change of coordination number at their substitution sites are shown different from others using OFM descriptors. The negative-formation-energy-family SmFe12－α－β[Al/Ga]αYβ structures show a common trend of increasing coordination number at substituted sites, whereas structures with positive formation energy show a corresponding decreasing trend.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
We established a query-and-learn framework including 1-new structure generator using substitution method, 2-machine learning model to suggest new calculated datapoint and 3-first-principle calculation to validate data suggestion. We also built components to monitor the structure discovery as 1-embedding representation show information of calculated/non-calculated structures, 2-Bhattacharyya coefficient to measure co-existence between representation features and expected properties. All these components show our framework as an interpretable machine learning method for structure discovery.
|
| Strategy for Future Research Activity |
We extend our research into a larger structure screening space with the decomposition phenomena is reproducible using the structure generator and the proposed query-and-learn framework shown as an inference to unveil meaningful correlations. In nature, SmFe12-based structures often decompose into multiple sub-phase compounds rather than single crystal magnets in the bulk form. Decompose mechanism relies strongly on the stability of compounds synthetic from any pairwise of substituted elements, Fe and Sm, with arbitrary elemental ratios. We expect that monitoring time-series data of structures prawn from genetic-algorithm generator e.g., USPEX using features of the query-and-learn framework, will bring meaningful insights to decomposition mechanisms in this structure family.
|
| Causes of Carryover |
We plan to build a bigger calculation system in the next fiscal year. Therefore, all the remaining budget this year will be merged with budget in the next year to use for this purpose.
|
