Research On Performance Of Perovskite Materials Via Machine Learning

In materials discovery,it is key to explore the structure-composition-property relationships,and machine learning can be used as an effective tool.However,the complexity of conventional machine-learning and the lack of model interpretability make it difficult to derive simple descriptive formulas.Perovskite materials have received much attention due to their exceedingly good performance in fuel cells and electrocatalysis.Thermodynamic stability is a key parameter that broadly determines whether the material is expected to be synthesized and whether it will decompose under certain operating conditions.The thermal and chemical stability,to a large extent,depends on the formation energy.Band gap is an important parameter that determines the optical properties of perovskite materials and controls the performance of various optoelectronic devices.In high-temperature fuel cell technology,oxygen vacancy formation energy is an important indicator to achieve rapid oxygen diffusion and oxygen catalysis.High throughput calculations based on first principles(DFT)to obtain formation energy,stability,band gap and oxygen vacancy formation energy with high accuracy require a lot of time and effort and are inefficient,and it is also impractical to obtain formation energy,stability,band gap and oxygen vacancy formation energy for a large number of perovskite material systems through experiments.So the artificial intelligence based machine learning prediction method becomes an efficient alternative method.The main contents of this paper include the following.In the first part,we propose a new method that combines extreme feature engineering and linear regression to explore the structure-composition-property relationships of perovskite materials.The relationship between the thermodynamic stability and lattice constant of perovskite and other features is discussed.A large number of new descriptors were constructed by using extreme feature engineering,and important subsets of descriptors were obtained through feature selection,and then the optimal descriptors were found through linear regression algorithm,and new expressions of thermodynamic stability and lattice constant were obtained.In the second part,features are compared and valid features are selected as descriptors to build prediction models based on different machine learning algorithms such as random forest(RF),gradient boosted regression tree(GBR),adaptive boosting algorithm(AdaBoost),categorical gradient boosted tree(Cat Boost)and extreme gradient boosted tree(XGBoost).The relationships with the formation energy,stability,band gap and oxygen vacancy formation energy of ABO₃ perovskite are investigated and models with high accuracy are obtained for the prediction of these four key properties.With the trained machine learning model,the key features affecting the perovskite properties are found.The results show that this method,which combines extreme feature engineering and linear regression,can explore the structure-composition-property relationships of materials without a priori knowledge,accelerating the development of materials,and providing a new method for materials exploration and research.The model has a very high accuracy in the study of perovskite property prediction and can be widely used in the study of perovskite materials that require high volume property prediction.