Material Composition Generation And Material Crystal Property Prediction Based On Deep Learning

The successful application of deep learning in natural language processing and computer vision has attracted the attention of many traditional fields.In recent years,with the development of digitalization in the field of materials science,deep learning has made breakthroughs in the application of material generation tasks and material crystal property prediction tasks.However,as far as current research is known,there are still some problems in the application of deep learning in material generation tasks and material crystal property prediction tasks,which affect the progress of new materials exploration.In view of the current problems,this paper proposes corresponding solutions:(1)Therefore,in order to improve the efficiency of new material exploration,this paper proposes a Conditional Generative Adversarial Autoencoder(Con ALAE)model to solve the three problems of deep generative methods.In response to the problem of insufficient novelty of the generated materials,Con ALAE uses the generative confrontation network(GAN)framework for material generation,and introduces the idea of variational autoencoders(VAE),which improves the problem of easy collapse of the GAN model and improves the stability of model training? In addition,in view of the problem that the model cannot generate materials with specific target properties,Con ALAE introduces conditional generation,which encodes material property information as model condition input,so that the model can control the generation of materials that meet specific target properties? Finally,for the limitation of specific material systems,the composition matrix that can uniquely represent the material composition of each series is used as the input data of Con ALAE,so that the model can be generated across material systems.In this paper,a series of experiments are performed on two large open materials datasets,OQMD(Open Quantum Materials Database)and MP(Materials Project),to verify the conditional generation ability of the Con ALAE model.(2)In order to improve the prediction accuracy of material crystal properties,this paper uses the idea of Graph transformer,comprehensively considers the two aspects of crystal geometric structure characteristics and node correlation,and proposes the Crystal Transformer method.The Crystal Transformer model introduces crystal geometric structure information into the input features as the global feature of the crystal graph,and uses the multi-head attention mechanism to calculate the correlation between nodes to aggregate the adjacent local features of the nodes in the material crystal graph.Thus,each update of the model realizes the information exchange from local to global in the crystal graph.In addition,in order to verify the effectiveness of the Crystal Transformer method,seven kinds of material property prediction experiments were carried out on the 60k+crystal material data set on the MP data set,and all of them achieved the current relatively advanced property prediction performance.