[22a-M206-1~11] 23.1 Joint Session N "Informatics"

[22a-M206-5] Supervised Learning of High-dimensional Output Variables in Materials Informatics

〇(D)Megumi Iwayama^1,2, Stephen Wu^1,3, Chang Liu³, Ryo Yoshida^1,3,4 (1.SOKENDAI, 2.Daicel Corp., 3.ISM, 4.NIMS)

Keywords:materials informatics, machine learning, high-dimensional output variables

Supervised learning problems in most previous studies, the output is a scalar or low-dimensional real-valued vector. On the other hand, there are many problem settings in materials research where the output variables such as functional data and electron microscope images are inherently ultra-high dimensional and the number of available data is limited, but the methodology of supervised learning in such cases has not been well studied. In this study, we report the research results of supervised learning of high-dimentional output variables by a generative adversarial networks and a uniquely modeled method based on functional data analysis.