Materials search method using high-throughput experimental screening and machine learning models

Abstract

To improve properties of materials, the effect of additive elements was investigated using high-throughput (HT) experiments and materials informatics (MI) technique. In case of MI, a small number of datasets becomes a problem. Furthermore, the data in the literatures were not obtained under the same conditions, so the homogeneity of the data is questionable. The use of such datasets makes it difficult to search for materials by MI. As a countermeasure, we should collect data under the same conditions by our own HT experiments. Especially, we would like to suggest that a large number of homogeneous datasets is better than a small number of highly accurate datasets for MI. In this study, we focused on phosphorescent materials [1], oxide ion conductors [2] and electrode materials for lithium-ion battery [3].

At first, material libraries were synthesized by HT experiments. In case of oxide materials, the ink-jet technique as shown in Fig. 1 [2, 3] is one of the synthetic method suitable for HT experiments, therefore we have mainly used the instrument, which had 4 ink-jet heads. In case of the ink-jet method, the so-called “coffee stain effect” often becomes a problem to cause inhomogeneous film quality. Multi-target sputtering equipment [1] is another suitable method, especially searching for alloys and semiconductors. However, the expensive target can be often problem.

Next, the properties of material libraries were estimated using HT measurement system, for example X-ray diffraction (XRD), X-ray fluorescent analysis (XRF) and X-ray absorption near edge structure (XANES) spectrum measured at SPring-8. Alternatively, we made our own dedicated HT measuring device. In this case, it is important to screen the materials using properties that can measure at high-speed without laborious experiments. Data from such HT experiments are guaranteed to be homogeneous and suitable for material screening. Note that the purpose of this study is the screening for materials, not a prediction.

Appropriate materials were searched by machine learning models using composition-based explanatory variables and experimentally-obtained objective variables. In many cases, we trained a machine learning model using the dataset by leave-one-out cross validation (LOOCV) method because the dataset is not big data. The regression model selection methods are case-by-case and there is no universal selection method.

Finally, specimens proposed by MI were synthesized, often by solid-state reaction method, and then the properties of the materials were verified experimentally. In conclusion, the materials with high properties could be searched in a shorter time. The results suggest that the combination between the HT experiments and the MI technique is effective for searching additives under the limited conditions like this study.

However, MI did not show the reason for the selection of the materials. It is very difficult to explain the reason why the MI choose the materials because the machine learning system is “black box”. If there is big data about the materials, the problem will be clarified. In the next project, we will investigate the reason and the chemical interpretability of the regression results.

[1] H. Hazama et al., Inorg. Chem., 58 (2019) 10936-10943, [2] M. Matsubara et al., ACS Comb. Sci., 21(2019) 400-407. [3] S. Tajima et al., STAM method, (submitted).