Data-Driven Crystal Growth Using Flux-Method Process Informatics

Crystallographic characteristics such as crystallinity, crystal outline, and size are one of important factors to determine material performance, since they relates to physicochemical phenomena occurring on the surface and inside of materials. The flux method, which grow crystals in molten salts (fluxes) is a powerful technique for development of high-performance crystalline materials. However, its growth guidelines are not well established due to multi-step growth process with various phenomena. Therefore, it is necessary to explore the optimal conditions in a huge experimental space based on various experimental factors, including flux species. Thus, it takes several years to develop crystals using flux method.<br/>Recently, we have studies flux-method process informatics (FPI), which is a data-driven approach to effectively explore favorable experimental conditions based on crystal growth prediction in the flux method. There are many issues to be solved to achieve high-accuracy FPI system. For example, description manner of feature values, big-data collection, expansion of materials target, and also effectivity of this method itself. In this study, we constructed an adaptive design of experiments (ADOE) system that can be used for FPI and applied to an anisotropic perovskite-type oxides.<br/>The ADOE system accords to a Bayesian optimization cycle consisting of (I) acquisition of experimental data, (II) modeling by Gaussian process regression, (III) conduction of virtual experiments over 10000 ways, and (IV) proposal of experimental conditions based on the acquisition function. In this time, a layered perovskite oxide Ba₅Nb₄O₁₅ (BNO) with an anisotropic crystal structure was selected as one of the model materials. As the explanatory variables, experimental conditions, including raw material amounts, flux species, and heating conditions were used. Two types of crystal sizes were used as the objective variables to describe anisotropic crystal shape of BNO. The number of training data was about 70.<br/>Firstly, we evaluated the efficiency of FPI using already collected dataset. The all dataset was divided into 10 as a training data and others as a test data. A prediction model was created using the training data, and the test data with high-likelihood one was selected as the goal candidate. If the selected test data was different from the goal, the data was added to the training data and continued the exploration until finding the goal. The cycles of modeling were output per 1 trial and scored the efficiency of ADOE after 100 trials. Comparing 100 random experiments, the average number of cycles by ADOE was 6 times smaller, indicating the better efficiency of this system than human-driven approach. In this presentation, we will also discuss the experimental results after applying the ADOE system, and contributing mechanism of each factor on the crystal growth.<br/><br/><b>Acknowledgement</b><b>s</b><br/>This research was partially supported by NEDO Feasibility Study Program, NEDO Green Innovation fund projects, JST Open Innovation Platform with Enterprises, Research Institutes, and Academia (JPMJOP1843), JSPS KAKENHI (Grant Number 21K04807, 22H04533, and 22H00568), Knowledge Hub Aichi, TAKEUCHI Scholarship Foundation, and Wakasato association in Shinshu university.