Editor’s Note:
Recently, a new active learning material design framework developed by Prof. Wang Jinlan from School of Physics has achieved fast and accurate prediction of 2D ferromagnetic materials with high Curie temperature, breaking the bottleneck of searching complex materials and property space. Related results were published in Chem (sister journal of Cell) titled “On-the-fly Interpretable Machine Learning for Rapid Discovery of Two-dimensional Ferromagnets with High Curie Temperature”.
In recent years, AI technology has set off a wave of technological change in the field of materials science. The combination of advanced machine learning techniques with traditional experimental or theoretical calculations has not only significantly accelerated the material development, but also provided new insights into material structure-activity relationship. Although this brand-new scientific research paradigm has made remarkable achievements, its application in complex systems and properties is still in its infancy and still faces many difficulties and challenges.
The 2D materials feature diverse crystal structures with complex factors affecting the ferromagnetic properties (structure, composition, electron, spin, etc.). Therefore, it is difficult to apply machine learning to 2D magnetic systems. Previously, the research team has found certain number of ferromagnetic materials from the existing 2D materials database (Adv. Mater. 2020, 32, 2002658) by means of machine learning and the first-principles high-throughput computation. But the actual scale of possible chemical space far exceeds the existing databases. In a larger range of chemical space, it is more likely to find excellent 2D ferromagnetic materials. However, the lack of available data, the inefficiency of sampling functions, the shortage of deep physical insights in descriptors, and the lack of interpretability of machine learning models all restrict the search of magnetic chemical space.
In order to solve this problem regularly available in the field of machine learning materials design, Prof. Wang Jinlan's research team proposed an active learning framework with feedback iteration function, which combined integrated learning algorithm and edge sampling algorithm to explore the magnetic chemistry space adaptively. This active learning framework integrates feature engineering, model learning, data sampling, first-principles calculation and model interpretation, and can efficiently explore the chemical space in the case of lack of data and high feature dimensions. Finally, 9622 2D ferromagnetic candidate materials were screened from 200,000 candidate compounds, among which 722 were 2D ferromagnetic materials with high Curie temperature. This work breaks through the bottleneck of machine learning technology in the application of complex systems and properties, and provides a new strategy with great potential for rapid and accurate exploration of large chemical space.
Dr. Lu Shuaihua and Associate Professor Zhou Qionghua from School of Physics, SEU are the first authors, with Prof. Wang Jinlan as the only corresponding author and SEU as the only completion institute of this work. This work was supported by the National Key Research and Development Program, the National Natural Science Foundation of China, the Fundamental Research Funds for the Central Universities and the Natural Science Foundation of Jiangsu Province.
Paper’s link:
https://www.cell.com/chem/fulltext/S2451-9294 (21), 00580-5
Submitted by: School of Physics
Translated by Melody Zhang
Edited by Pang Xuan, Qi Yuchen
For reprint or any commercial/public use, please contact us.
