4-7. XenonPy: a machine learning platform for accelerating materials design

Chang. Liu

The Institute of Statistical Mathematics, Tachikawa, JP

Abstract: In recent years, various kinds of machine learning (ML) algorithms have been applied to material science and gained lots of outstanding successes. We believe that these increasing successes finally will bring the materials science to an entirely new stage. However, using cutting-edge methodologies in the research is also associated with risks. Firstly, almost all of these ML methods are implemented in different packages, and also the authors are not professional software engineers. Additionally, in most case, researchers built these packages just for the paper submission, which means that the packages will never be maintained after the paper publication. The risk of stability has become to be a barrier to the growth of material science. We are building an all-in-one materials informatics platform called XenonPy to address these issues. Different from other packages, we focus on the application of material informatics with long-time-support. Currently, this platform comes with a same-named package and a pre-trained model database. For the package, called XenonPy¹, we selected the most trustable packages, e.g., pymatgen, RDKit, and PyTorch, as our foundation packages and upper them, we built a flexible, easy-to-use, and extensible interface to connect every part. The pre-trained model library, named XenonPy.MDL², which can be used to predict various properties of small molecules, polymers, and inorganic solid-state materials. Along with this database, we demonstrated the outstanding successful applications of transfer learning in our present paper.