2018 Fiscal Year Research-status Report
Inverse materials design by integrating transfer learning techniques into a Bayesian framework
| Project/Area Number |
18K18017
|
|---|
| 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, 助教 (70804186)
|
|---|
| Project Period (FY) |
2018-04-01 – 2021-03-31
|
| Keywords | Transfer learning / Materials informatics / Polymer design / Open source software |
|---|
| Outline of Annual Research Achievements |
In 2018, I have successfully completed an open Python package, called XenonPy, that can perform Bayesian inverse design for organic molecules, as originally planned in my proposal. Moreover, a case study of designing high thermal conductivity polymer was completed, with a peer-reviewed paper conditionally accepted by npj Computational Materials under minor revision. One transfer learning algorithm, called Frozen-featurizer, was also implemented in XenonPy. I am now starting a series of simulation and experiment regarding the use of this algorithm. Meanwhile, more variety of the transfer learning algorithm will be developed and analyzed to understand the best way to exploit transfer learning for material design with small data.
|
| Current Status of Research Progress |
Current Status of Research Progress
1: Research has progressed more than it was originally planned.
Reason
The basic functionality of the open software for material design, including the Bayesian inverse design and one transfer learning method, was finished as planned. One paper is conditionally accepted along with a patent submitted for approval. Furthermore, the developed methodology on transfer learning has been extended to other applications outside of materials science as well, showing the great potential of our algorithm. I have given a presentation on this subject in four Japanese conferences and one international conference, as well as being an invited speaker at four international workshops. Despite the great success on the software development and application, I think the experimental work in this project needs to be improved in the next year.
|
| Strategy for Future Research Activity |
With the basic infrastructure (the open package) completely ready for sure, I will continue on the development and experiment of transfer learning algorithms, as planned in the proposal. Furthermore, more collaboration with the experimental group at Tokyo Institute of Technology is expected to speed up the experimental progress. At least one more paper is expected to be published next year along with 3 or more conference participations. The major challenge is expected to the on the generalization and automation of the transfer learning algorithm to efficiently produce reliable predictive models for thermal conductivity of polymers with limited data. I expect that more effort will be spent on understanding the underlying mechanism of successful knowledge transfer in the next year.
|
| Causes of Carryover |
N/A
|
