Accelerated Discovery of Novel Inorganic Materials with Desired Properties Using Active Learning

Abstract

Construction of prediction models using machine learning algorithms on existing databases expands the search limit of undiscovered structures, in principle, to the entire materials space. However, because of uncertainties in machine learning prediction, the suggested properties are not always promising; thus, improving the database quality is mandatory for validation as well as improvement in prediction accuracy. To achieve this, we herein implement an active learning process, beginning with a limited number of databases, to find materials satisfying target properties (band gap and refractive index) with minimized trials and errors. The regression model is initially trained with only around 2% of the entire search space, and 20 new databases, suggested from the optimization schemes, are added at each optimization process. Between exploration, exploitation, random selection, and the Bayesian optimization method, the Bayesian method exhibits the best performance in finding the number of materials that satisfies the criteria within limited trials In addition, the structure with the maximum target property values is found after searching only around 7.0% and 7.7% of the entire database for band gap and refractive index, respectively. Current results clearly confirm that the active learning process can be accelerated to find ideal materials satisfying target properties with minimized resources.