| Thermoelectric material is a new type of energy material that can realize the direct mutual conversion of thermal energy and electrical energy.It has the advantages of safety,energy saving,environmental protection,etc.It is beneficial to the development of industry and has no pollution to the environment,so it has received extensive attention.However,the prediction of novel thermoelectric materials by traditional experimental and computational methods is costly and time-consuming.In this paper,four key features that contribute most to thermoelectric performance are screened out by machine learning method,and the relationship between the key features and thermoelectric performance is found through model decomposition and feature combination analysis.In addition,we provide a machine learning method based on feature selection to efficiently and accurately predict the thermoelectric properties of130,000 materials.The main research contents are as follows:1.Machine learning method is adopted to screen out features that have great influence on thermoelectric performance.Data are obtained from the MRL(Materials Research Laboratory)database developed by UCSB.145 initial features were calculated and extracted from the chemical formulas of the materials using the Magpie package.Information entropy evaluation based on Extra Trees model ranked the importance of the initial features,removed the correlation and redundancy features,and screened the four key features that contributed most to the thermoelectric performance.We further explored the relationship between the four key features screened out and the thermoelectric performance.Firstly,we explored the distribution law of thermoelectric materials and non-thermoelectric materials in each feature,and found that the distribution law of thermoelectric materials and non-thermoelectric materials in each feature was not obvious,and there was a lot of overlap.Secondly,we explored the relationship between each feature and thermoelectric performance,and found that each feature has no obvious regularity and linear relationship with ZT value,electrical conductivity,thermal conductivity and Seebeck coefficient.Finally,we try to combine the features to explore the relationship between the combined effect of the features and the thermoelectric performance.It is found that the combined effect of features indirectly affects the thermoelectric performance by affecting the electrical conductivity or thermal conductivity.The research content not only provides important information for the further study of thermoelectric materials,but also provides some research ideas for the development of thermoelectric materials.2.Prediction of thermoelectric materials by machine learning optimal model based on feature selection.According to the comparison of Random Forest,Ada Boost,Support Vector Machine,Gradient Boost Decision Tree and K-Nearest Neighbor five models of accuracy,determination coefficient and test set prediction evaluation,it is found that the Random Forest model is the best comprehensive model among the five models.Based on the selected four features that contribute most to thermoelectric performance,the Random Forest model was selected to predict and analyze thermoelectric materials,and reduce the search space of more than 130,000 materials to 6476 material candidate systems,among which 138 thermoelectric candidate materials have a probability of more than 95%.Among the 138 predicted thermoelectric materials,Ag-containing materials occupy more than 40.Therefore,we further explored the relationship between Ag and thermoelectric performance,and found that the spatial structure,Ag content and unit cell volume of Ag-containing materials affect the thermoelectric properties.The method adopted in this paper is not limited to the search for new thermoelectric materials,but can also be applied to search for other functional materials. |
