Phase Diagram Construction by Active Learning

Ryo Tamura^1* and Kei Terayama^2*

¹ International Center for Materials Nanoarchitectonics (WPI-MANA), National Institute for Materials Science, Tsukuba 305-0044, Japan

² Graduate School of Medical Life Science, Yokohama City University, Yokohama 230-0045, Japan

ABSTRACT: In data-driven materials discovery, Bayesian optimization has been attracted attention, which is a black-box optimization method using Gaussian process. In Bayesian optimization, based on the machine learning prediction, materials candidates which would have desired properties to be synthesized/measured in the next step are efficiently selected. On the other hand, in general, before conducting property optimization, there is a need to draw a phase diagram of target materials. Since this problem cannot be directly performed by Bayesian optimization, it is necessary to develop a data-driven method suitable for phase diagram construction. We have developed a method realizing an efficient phase diagram construction by active learning called PDC (Phase Diagram Construction). In this method, the most uncertain point in the phase diagram is selected as a candidate to be synthesized/measured at the next step by using uncertainty sampling. Then, an experiment is performed according to the selected candidate and the phase at the point is identified. Uncertainty sampling is performed again on the increased experimental data to select next candidates. By repeating this, it is possible to quickly draw a phase diagram (Fig. 1). By using PDC, experimental determination of an uninvestigated phase diagram for the deposition of Zn–Sn–P films by molecular beam epitaxy is attempted.

 Figure 1.  Flow of phase diagram construction by active learning.

Keywords：phase diagram; active learning; uncertainty sampling