Efficient Screening Of Perovskite Oxide Photocatalysts For Overall Water Splitting Via Machine Learning

In order to solve the increasingly severe energy and environmental problems,the photocatalytic overall water splitting method has been raised.The core of the method is to find high-efficiency photocatalysts whose energy band structures satisfy the photocatalytic conditions.Perovskite oxide materials have attracted extensive attention due to their stable structures and excellent properties.The various combinations of elements make perovskite oxides rich in species,making it possible to find suitable energy band structures for photocatalytic materials.However,it costs too much if we use traditional trial-and-error screening method for such huge combination conditions.In recent years,machine learning method has been introduced into the design of new materials.As a data-driven method,it enables the efficient screening of a large number of materials.In this paper,we aimed at the problem of how to speed up the search for photocatalytic oxide materials for overall water splitting.We combined machine learning algorithms with first-principles calculations and predict over 8,000photocatalyst materials for overall water splitting in the visible light range from more than 60,000 perovskite materials.On the one hand,some catalyst candidate materials were selected for experimental synthesis.And on the other hand,the structure-activity relationship of the B-site element of perovskite materials was summarized to further accelerate the selection of single and double perovskite oxides.The main contents include the following two aspects:1)Machine learning accelerated screening of single perovskite oxides.Aiming at the problem that there are many combinations of perovskite oxide elements and the difficulties in high-throughput computational screening,we use machine learning methods to perform feature construction and feature engineering screening of atomic information and structural information.Through two-step modeling,classification and regression models were established successively,and 350 candidate materials with band gaps suitable for photocatalysis are predicted from 2731 single perovskite materials,of which 151 materials do not contain lanthanides element.On the one hand,this work confirms the feasibility and accuracy of the machine learning model for predicting the band gap of perovskite oxides,and on the other hand,151 photocatalyst candidate materials in the visible light range are screened.This part of the work has greatly narrowed the scope of material selection.2)Machine learning accelerated screening and structure-activity relationship analysis of double perovskite oxides.Compared with single perovskite oxides,A₂BB’O₆ type double perovskite oxide materials have more diverse combinations of elements and a wider range of energy band regulation.As a result,double perovskite oxides are better candidates for photocatalysis.But more element combinations make it more difficult to screen out candidate materials.In this section,we construct a complete training data set,select suitable feature descriptors,and train an accurate machine learning model.On the one hand,we have screened nearly 8,000 possible photocatalytic overall water splitting materials from more than 50,000 double perovskite materials that have not been included in the database.On the other hand,we are more focused on analyzing the statistical laws of materials.We discuss how to pick the best combination from a multitude of element combinations from the top design perspective.This part of the work not only narrows the selection range of double perovskite materials,but also provides regular guidance for the search for new double perovskite oxide photocatalysts.