Research On The Performance Of Perovskite Photocatalyst Based On Machine Learning

The development of society has an increasing demand for functional materials.A great deal of experience and data has been accumulated in the research of functional materials for a long time.In recent years,with the popularization of the Internet and computer technology,many databases have been established on the properties,structures and applications of materials.How to use these data to assist the discovery of new materials has become a hot topic in material research.ABO₃perovskites are considered as one of the most promising photocatalytic materials benefiting from their excellent structural flexibility,good stability,tunable bandgap（E_g）,high photocatalytic activity,low price,and simple synthesis method.However,the existing ABO₃perovskites have three main blemishes:1)a large E_glimits sunlight from being absorbed,2)a high rate of carrier recombination leads to decreased catalytic activity,and 3)a low specific surface area（SSA）limits the reactant to adsorb on the surface.Therefore,it is necessary and meaningful to develop a multi-objective stepwise design strategy based on machine learning（ML）to design ABO₃perovskite candidate samples and accelerate the discovery of perovskite comprehensively.The main research objective of this paper is to use ML technology to establish the quantitative structure-property relationship between the E_g,SSA,and crystallite size（CS）of ABO₃perovskites and their atomic parameters,chemical composition,and experimental conditions,thereby accelerating the design and development of new ABO₃perovskite materials.Furthermore,the influence of experimental conditions on these properties is also analyzed to study their formation trends,thus achieving a more accurate and reasonable material design.The main contents of the paper include the following aspects:（1）Introduce the concept of ML and material design,the general process of ML in material design.It also introduces the basic principles of photocatalysis,the structure of ABO₃perovskite and its application in photocatalysis,and the application and development of ML in the field of ABO₃perovskite photocatalysis.（2）The ML models for the E_g,SSA,and CS of the ABO₃perovskite material are established.The gradient boosting regression（GBR）,support vector machine regression（SVR）,backpropagation artificial neural network（BPANN）and multiple linear regression（MLR）algorithms are compared in the modelling of ABO₃perovskite properties.The results show that the GBR algorithm is more suitable for the modelling of E_gand CS of ABO₃perovskite materials,while the SVR regression is more appropriate for the modelling of their SSA.After feature screening,the optimal feature subset of the E_gof ABO₃perovskite materials contains six features.The boiling point of the B-site in these six features has the largest influence on the model and is a positive correlation with their E_g.The best feature set of SSA of ABO₃perovskite materials contains ten feature variables,among which the calcination temperature（CTP）and the electron affinity of the B-site play an important role in the model.The CS of ABO₃perovskite materials contains nine feature variables,among which the importance of the preparation method（PM）and CTP is very high.After the hyperparameter optimization of the GBR model of the E_gand CS,the correlation coefficient（R）values of the leave-one-out cross-validation（LOOCV）and the testing set are 0.9213,0.8976 for E_gand 0.9018,0.8778 for CS,respectively.And after the hyperparameter optimization of the SVR model of the SSA,the R of the LOOCV and the testing set are 0.8915 and 0.8480,respectively.In addition,these models have also been developed as online prediction applications for the experimental researchers interested in the applications of models.Through stepwise screening,35 ABO₃perovskite candidate materials with appropriate E_gvalues（1.4～2.6 e V）,high SSA(>60 m²g^-1),and small CS（<36 nm）were screened out from5368 candidate ABO₃compounds.Furthermore,a ML model for predicting the hydrogen production rate(R_H2)of photocatalytic water splitting was established to predict the R_H2of the 35 proposed ABO₃perovskite candidate materials.The prediction results show that the R_H2of these candidate materials are all greater than6000μmol h^-1g^-1,which indicates that they have great potential in the application of photocatalisis.（3）The influence of the PM,CTP,and calcination time（CT）on the E_g,SSA,and CS is analyzed.It is found that these experimental conditions have no obvious effect on the E_g.That indicates the E_gmay depend more on the elemental composition of the material.The suitable E_gis more likely to acquire when the main elements at the A site are Bi,La,Pr,and at the B site are Fe,Ti,Mn.The SSA and CS are more easily affected by experimental conditions,especially the CTP.The higher the CTP,the lower the SSA and the larger the CS.Besides,based on the prediction results of ML,we have also analyzed the relationship between these three properties and the R_H2.The results showed that the perovskites with a E_gvalue in the range of 1.6～2.4e V and a high SSA have a higher R_H2.Although the effect of CS on R_H2is not so obvious,it also shows that smaller CS is more inclined to obtain a higher R_H2.