| Novel functional materials discovery is of great significance to promoting social development,especially the advancement of industrial technology.It is aimed at efficient inverse design of materials,and the process requires efficient methods to explore the broad chemical space of target regions,as well as fast and accurate methods to predict the properties of candidate materials.At the same time,due to the extensive existence of crystalline materials and data-driven requirements,inorganic crystal structure prediction(CSP)is one of the most basic tasks in materials discovery,and it is also the focus of this work.Traditional reverse design strategies mainly include Highthroughput Virtual Screening(HTVS)and Global Optimization(GO),but they have limitations such as consuming a lot of resources for Density Functional Theory(DFT)calculations.With the development of deep learning,the application of Generative Models(GMs)to crystal structure prediction is considered to be a potential new method.But it has the problems that the reversible representation of the periodic crystal structure is difficult,the reversible representation needs to meet the requirements of symmetry invariance such as translation and rotation,and the structural diversity of each element of the inorganic crystal is low and unbalanced.Aiming at the problems existing in the traditional generative model in CSP,this work proposes two different solutions which greatly improve the performance.The research contents and results are as follows:1.Crystals Wasserstein GAN(CrystalsWGAN),which is a crystal structure prediction model based on generative adversarial networks:having abandoned the traditional voxelized representation,using point cloud representation based on atomic coordinates and unit cell parameters,and using data augmentation strategies such as generating supercells,translations,flips and shuffling the order of points.In this work,we build a reverse design framework based on Wasserstein GAN with Gradient Penalty(WGANGP).Referring to the interaction between particles,we make the discriminator extract the local and global features of the unit cell structure.This proposed model is applied to the binary V-O system,and a new V-O training set is constructed by replacing the binary compounds in the Materials Project.The V-O crystal structures are generated using the converged model after training,and screened by chemical formula and interatomic distance.Then preliminary analysis of chemical compositions and synthesizability was performed,followed by DFT calculations,The thermodynamic stability of the optimized structures can be judged by the formation energy,and new stable materials can be finally screened out.Experiments show that the synthesizability of the materials generated by the model is as high as 41.73%.After 1132 structures were optimized by DFT,215 structures that satisfy the thermodynamic stability conditions were finally found,including 48 structures with new compositions beyond the MP database,of which 10 materials are considered synthesizable.2.Energy-constrained crystals Wasserstein GAN(ECCWGAN),which is a crystal structure prediction model based on generative adversarial networks with energy constraints:the essence of the inverse design of inorganic materials is to achieve the inverse mapping of new materials with target properties,so it is necessary to optimize the properties of latent space.This work proposes an energy-constrained crystal structure prediction model,which is still based on WGAN-GP and consists of three network components:generator,discriminator,and predictor.Referring to Conditional Generative Adversarial Nets(CGAN),this work introduces additional label information,and also introduces additional loss to the objective function through a feedback loop(called a predictor)to train the generator for automatic optimization,which also enables the model to generate stable structures more efficiently.The proposed model is applied to the binary V-O system.We verify the structural validity and synthesizability,and DFT calculation is performed to screen the final stable structures.Experiments show that the synthesizability of the materials generated by the model is as high as 55.61%.After 879 structures were optimized by DFT,224 structures that satisfy the thermodynamic stability conditions were finally found,including 57 structures with new compositions beyond the MP database,of which 14 materials are considered synthesizable.In this thesis,two prediction framework s based on generative adversarial networks are constructed for inorganic crystal structure prediction.We use them for binary V-O systems.The experimental results show that the two methods proposed in this thesis can generate a large number of stable new V-O materials.Both of methods are successful inverse design strategies.They can not only overcome the drawbacks of traditional HTVS methods and GO-based strategies,but also propose new solutions for the difficulties and deficiencies in the ap plication of current generative models in materials design.The idea is of great significance for accelerating the inverse design of inorganic materials. |
