| 研究実績の概要 |
In 2019, I have successfully upgraded the open Python package, called XenonPy, that can perform Bayesian inverse design for organic molecules (implementation completed in 2018). The upgrades were published in Molecular Informatics. One of the upgrades is the improvement of transfer learning efficiency of the Frozen-featurizer function in XenonPy. As a result, more than 140,000 pre-trained predictive models for various material properties are now openly available in our model library server and the details are explained in our publication in ACS Central Science. Using these models, second round of high thermal conductivity polymer design has started and the experimental work will be performed next month (May 2020). Extra transfer learning studies on different materials, such as inorganic materials, have also been started to provides hints for understanding the mechanics of transfer learning.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
1: 当初の計画以上に進展している
理由
The functionality of the open software for material design, including the Bayesian inverse design and transfer learning methods, were further improved for a broader range of applications. Three papers have been published, which is more than originally planned. The progress of this project has motivated a new bigger project that will focus on the development of an open polymer database in Japan. Not only a second round of experiment will be performed, but also the design target has been extended from polymers built with one kind of monomer to polymers built with multiple monomers and additional processing methods during synthesis of the polymers. The coverage of the domain of polymer design has become larger than originally planned using my materials informatics tools.
|
| 今後の研究の推進方策 |
With the improved infrastructure (new version of XenonPy) in place, I will consider extending the software to cover analysis tools for transfer learning in materials informatics. So far, there has not been much progress on the understanding of the underlying mechanism of successful knowledge transfer between different material properties. While the experimental works are progressing smoothly as planned, I would like to challenge myself to build new statistical methods for the transfer learning framework. I will begin with new case studies other than thermal conductivity of polymers to look for hints of common patterns in successful transfer learning cases. Hypothesis will be first tested using conventional statistical analysis tools.
|
