Searching High Spin Polarization Ferromagnet In Quaternary Heusler Alloy Via Machine Learning

In the study of spintronics,it has been expected to find more materials with high spin polarization and high Curie temperature above room temperature to ensure their daily use.At present,machine learning is becoming an important method to help material design and mining,which has been widely valued and applied.Applying machine learning to the research of spintronic materials will help to speed up the pace of research.In this dissertation,machine learning method combined with high throughput first-principle calculations is used to find stable materials with high spin polarization and high Curie temperature in Quaternary Heusler alloy.We use about 43000 materials collected in the open quantum materials database to train,verify and test the machine learning model to find the relationship between structure and stability.In the study,machine learning methods such as nearest neighbor regression,decision tree regression,random forest regression and deep neural network are used to statistically evaluate the performance of these models by using a variety of indicators(such as Pearson correlation coefficient,Determination coefficient and Mean absolute error).After comparison,it is found that the deep neural network model obtains the Pearson correlation coefficient of 0.99,and the Mean absolute error is only 0.021 e V / atom.On this basis,the machine learning method based on deep neural network model is finally used for research.24480 candidate materials were screened by using the deep neural network model with the best performance,and 4651 quaternary Heusler materials that may be synthesized and stable were obtained.Then,the first-principle calculation method based on density functional theory is used to calculate the high-throughput of these 4651 materials,and 189 possible high spin polarizability materials are found,including 149 half-metals,15 magnetic semiconductors and 25spin-gapless semiconductors.Further,the Curie temperature is estimated by mean field theory.It is found that 144 of these materials are higher than or close to room temperature.Finally,the phonon spectra of 16 kinds of magnetic half-metals and magnetic semiconductors with the highest temperature are further studied.It is found that these spintronic materials are dynamically stable.These results show that the workflow we designed is feasible and can make full use of the existing open database resources and use machine learning to do data mining,so as to carry out the research of spintronic materials efficiently and at low cost,and help improve the design and development efficiency of spintronic materials.