Research On New Material Discovery Algorithms For Inorganic Compounds Based On Deep Neural Networks

Materials are the material basis of human production and life,and the driving force that directly promotes social development.However,the current material discovery still involves major repeated experiments,which may take decades of research to determine materials suitable for technical applications.The main reason for this long discovery process is that the number of potential materials is huge,and it is thorny to choose the materials to be concerned about and which experiments to carry out.In the face of competitive manufacturing and rapid economic development,materials scientists and engineers must shorten the research and development cycle of new material discovery in order to solve the huge challenges of the 21 st century.However,the current research and development of new materials is mainly based on the scientific instincts of researchers and a large number of repeated "trial method" experiments.The number of materials simulated by experiments or high-performance principles is dwarfed by the expected potential diversity.In recent years,artificial intelligence(AI)has made exciting progress,among which the application of machine learning(ML)and deep learning(DL)technology has brought competitive performance to people in various fields.In addition,the availability of more and more material databases with experimental and/or computational characteristics has inspired researchers to adopt advanced data-driven material discovery technology interest in accelerating the discovery of new materials with selected engineering.Under the support of the National Natural Science Foundation project "Research on highthroughput composite material experimental phase diagram generation and phase identification method based on machine learning and image processing algorithm"(Project No.: 51741101),aiming at the problem of huge potential material quantity and unable to choose the material to be concerned wisely,this study adopts the deep learning method to design a generation model that can efficiently sample from inorganic compounds and a screening model that can accurately predict the properties of materials.Based on these two models,new materials for inorganic compounds are carried out.The discovery has important theoretical and practical significance for improving the research and development efficiency of materials.The main work and innovations are as follows:A GAN-based generation model is proposed,which can effectively sample from the huge chemical design space of inorganic materials.System experiments and verifications show that our GAN model can achieve a high degree of uniqueness,effectiveness and diversity in terms of generating capabilities.By expanding ICDS,Materials Project(MP)and OQMD,our generative model can be used to explore the unknown inorganic material design space.Compared with the thorough screening of billions of candidates,the derived extended database can be used for more efficient high-throughput computational screening.Although the principles of charge neutrality and electronegativity balance have been used to filter chemically incredible components in order to search new materials more efficiently,such clear composition rules are still too lax to ensure that Effective sampling of new materials.Although hypothetical materials with less than 5 elements can be cited(32 billion for 4-element materials with charge neutrality and balanced electronegativity),the design space for more elements can be challenging,and Mat GAN model can help a lot.A hierarchical characterization method for materials is designed.Based on the characterization method,a material property prediction model for convolutional neural networks with fully convolutional layers is proposed.By designing a special convolution operator,the model can extract useful features from the original input matrix of the material and perform the final regression task.Using the data in ICSD for supervised training of different tasks,a high-precision prediction model of ICSD-BG that can predict the band gap of the material and ICSD-FE that can predict the formation energy of the material is established.The prediction performance of the established model on the test set is basically consistent with the performance on the validation set,which ensures the reliability of subsequent screening of hypothetical materials generated by GAN with the prediction model.ICSD-FE is applied to the hypothetical materials generated by GAN-ICSD for screening,and materials with low formation energy are screened out.On the basis of screening materials with low formation energy,ICSD-BG was used to continue screening materials with a band gap between 1.0 and 2.0 e V.The band gap of the selected materials is 1.0 ～ 2.0e V and the formation energy of these materials predicted by ICSD-FE is placed on http: //github/danyabo/appendix.com for researchers to perform subsequent DFT simulation calculations or Experimental synthesis.