Identification of dominant factors contributing to photocurrent density of BiVO4 photoanodes using Machine learning

• Published in: Journal of photochemistry and photobiology. A, Chemistry. Chemistry., 2023-06, Vol.440, p.114651, Article 114651
• Main Authors: Kobayashi, Kazuki, Nagai, Yuya, Pan, Zhenhua, Katayama, Kenji
• Format: Article
• Language: English
• Subjects: Categorization; Descriptor selection; Machine learning; PEC performance; Photoelectrode
• ISSN: 1010-6030; 1873-2666
• DOI: 10.1016/j.jphotochem.2023.114651

[Display omitted] •Analytical data of photoanodes directly predict PEC performance.•Various machine learning method was used for the prediction.•The category of the JV curves were also predicted.•About a hundred samples of BiVO4 photoanodes were used.•Physical/chemical origins were identified. Bismuth vanadate (BiVO4) is one of the most promising materials for photoanodes in photoelectrochemical (PEC) water splitting. The PEC performance (measured by the current–voltage curve) is generally improved by the trial-and-error of the fabrication procedure; however the origin of the structural and physical properties is not always clear, especially in the developing stage. To clarify the physical origins for determination of the PEC performance, we used two types of the machine learning (ML) calculations for a combination of various analytical data of BiVO4 electrodes. First, the photocurrent density of the BiVO4 electrodes was predicted by the descriptors in the analytical data and the dominant factors were identified. The selected descriptors could provide a prediction of photocurrent values with the determination coefficient, ∼0.8. Secondly, we could predict two different types of the current–voltage curve by the ML categorization functions. Three dominant descriptors were identified and its accuracy was 100 %. The analysis of the dominant factors provided us useful information of the structural and physical properties to understand the photocurrent behavior without any prior knowledge.