Machine learning aided design of perovskite oxide materials for photocatalytic water splitting

Abstract

Suffering from the inefficient traditional trial-and-error methods and the huge searching space filled by millions of candidates, discovering new perovskite visible photocatalysts with higher hydrogen production rate (RH2) still remains a challenge in the field of photocatalytic water splitting (PWS). Herein, we established structural-property models targeted to RH2 and the proper bandgap (Eg) via machine learning (ML) technology to accelerate the discovery of efficient perovskite photocatalysts for PWS. The Pearson correlation coefficients (R) of leave-one-out cross validation (LOOCV) were adopted to compare the performances of different algorithms including gradient boosting regression (GBR), support vector regression (SVR), backpropagation artificial neural network (BPANN), and random forest (RF). It was found that the BPANN model showed the highest R values from LOOCV and testing data of 0.9897 and 0.9740 for RH2, while the GBR model had the best values of 0.9290 and 0.9207 for Eg. Furtherly, 14 potential PWS perovskite candidates were screened out from 30,000 ABO3-type perovskite structures under the criteria of structural stability, Eg, conduction band energy, valence band energy and RH2. The average RH2 of these 14 perovskites is 6.4% higher than the highest value in the training data set. Moreover, the online web servers were developed to share our prediction models, which could be accessible in http://materials-data-mining.com/ocpmdm/material_api/ahfga3d9puqlknig (Eg prediction) and http://materials-data-mining.com/ocpmdm/material_api/i0ucuyn3wsd14940 (RH2 prediction).