Automated stopping of spectral measurements with active learning

Abstract

There is a need for improvement of the efficiency, automation, and autonomy of various experiments in materials science with the recent rise of materials informatics. We have developed a method to improve the efficiency of spectral measurements, one of general experimental techniques in materials science, by using active learning. By using active learning to sequentially measure the energy points with the maximum of the acquisition function, we have achieved automated spectral measurement under optimal conditions without the intervention of an experimenter. By employing a stopping criterion based on the upper bound of the expected generalization error of the Gaussian process regression, the measurement can be automatically stopped regardless of the type of spectrum. This method allows us to obtain the materials information of equivalent quality with fewer measurement points compared to the conventional spectral measurement.