Prediction Of Superconductivity Of Materials Based On Machine Learning

Currently,the Fermi energy levels and the forbidden band widths of materials are usually obtained experimentally.There is no universally accepted theoretical method for determining the critical temperature of superconductors,which is usually studied empirically on a "trial and error" basis to discover new superconducting materials.However,both the experimental methods for obtaining the Fermi energy level and the forbidden band width and the empirical methods for discovering superconductors are very costly in terms of human and material resources.Therefore,it is urgent to explore new methods to obtain the properties of materials.As more and more materials are discovered and their data are more and more available,machine learning and neural networks with adaptive processing capability for large data provide new ideas for obtaining material properties.In this thesis,the identification of superconducting materials and the critical temperature of superconducting materials and the prediction of Fermi energy level and forbidden band width of materials are studied more deeply by machine learning methods,and the main research contents and conclusions are as follows:1)For the problem of superconducting material identification,this thesis introduces the XGBoost classification algorithm,using atomic frequencies as input,and the results show that the True Positive Rate(TPR)reaches 98.48%,which is better than the similar literature known to us so far.2)To address the problem of predicting the critical temperature of superconducting materials,a deep forest model is introduced in this thesis,using atomic frequencies as input,and the results show that the coefficient of determination,mean absolute error,and root mean square error on the test set reach 94.4%,4.04 K,and 7.51 K,respectively,all three of which are better than similar literature currently known to us.3)To address the problem of predicting the Fermi energy level of materials,a subnetwork model is proposed in this thesis,using atomic frequency,outermost electron number,electron affinity,first ionization energy,atomic number,and atomic radius as inputs,and the results show that the coefficient of determination,mean absolute error,and root mean square error reach 98.4%,0.10 eV,and 0.14 eV,respectively,on the test set,and all three indexes are better than what we currently known literature of the same kind.4)For the prediction of material forbidden band width prediction,a deep forest model is introduced in this thesis,using atomic frequencies as input,and the results show that the coefficient of determination,mean absolute error,and root mean square error in the test set reach 91.7%,0.27 eV,and 0.45 eV,respectively,and all three indexes are better than similar literature known to us so far.