Research On Thermoelectric Properties Prediction Of Materials Via Machine Learning

In recent years,thermoelectric materials have become a hot research direction in the field of materials because of their ability to directly convert between thermal energy and electrical energy,small size,high reliability and environmental friendliness.However,the research on thermoelectric materials,through a large number of experimental observations,and the traditional materials research methods based on intuition,have obviously not been applied to the rapid development of today's digital and intelligent era.Although we can use first principle to improve thermoelectric theory and design new materials,the research periods are still too long.Hence,it is of great significance to explore new research methods to accelerate the development of new thermoelectric materials.In the view of the problems mentioned above in material engineering applications and material design,this paper creatively proposes an prediction method of thermoelectric performance based on machine learning technology.We focus on establishing a model based on machine learning methods for predicting Seebeck coefficient of thermoelectric materials,so that,it can help material researchers screen for ideal thermoelectric materials.Besides,we use trained model to excavate important features of band structure having close relationship with Seebeck coefficient.Therefore,the work of this paper can roughly divided into two parts.For the first part,we obtain 22331 semiconductor materials Seebeck coefficient data record via Materials Project,an open source material database,and establish a data set for building regression models.Then,the feature engineering is carried out from two aspects of the constituent element properties and spatial structure information of the compound,and a "descriptor" is constructed by the selected element properties and spatial struceture for each compound in the data set.Then we use these descriptors to train the gradient boosting decision tree(GBDT)model.The predicted effect on the test set reached R2=0.81,MAE=20.52 ?V/K,RMSE=25.36 ?V/K,MAPE=5.62%.Finally,the importance degree of the input feature variables is given for the interpretability of the GBDT model.For the second part,in order to dig out the characteristics of the material band structure which are closely related to the Seebeck coefficient,we obtain 516 band structure data of semiconductor materials and use these data train a GBDT model which can predict Seebeck coefficient.Then we extract 4 important characteristics in band structure having close relationship with Seebeck coefficient with the help of this GBDT model and clustering.Finally,we verify the validity of the extracted characteristics.