Utilizing Neural Network Potential for Facilitating Experimental Discovery of New Metal Oxide Photocathodes Toward Water Splitting

Metal-oxide-based photocathodes are attractive components in photoelectrochemical cells for hydrogen production. This is due to both the ease of the synthesis and their stability under water-splitting conditions. However, most metal oxides are classified as photoanodes (namely, n-type semiconductors), resulting in a few metal-oxide-based photocathodes (p-type semiconductors). Therefore, in order to facilitate the experimental discovery of new metal oxide photocathodes, it is important to establish a methodology to easily classify candidates of the metal oxides as p- or n-types. In this study, we exhibit a new method to classify the p- and n-types. The key to developing our method lies in the combination of defect chemistry insight from materials science and the neural network potential of a machine learning technique. This combination has brought a sufficient classifier for semiconductor properties. Furthermore, the aforementioned method was employed to extract the candidates from the Materials Project database, leading to the experimental discovery of a new metal oxide photocathode.