Study On Material Structure And Properties Based On Machine Learning

Organic semiconductors,elastomers and other materials have a wide range of applications in light-emitting devices,rubber tires and other fields because of their superiority in mechanical and electrical properties.Factors affecting material properties include microstructure within the system,material system parameters,and material molecular structure.Traditional experimental studies of material properties mostly rely on trial and error and experience,which not only consume a lot of human and material resources,but also lack structural analysis and characterization of the system and molecules.Simulation methods such as molecular dynamics simulations to study material structure and properties require manual analysis,and the simulation runs for a long time and lacks the characterization of non-equilibrium processes.With the advancement of information technology,the application of machine learning algorithms in materials genetics has brought new life to the study of materials structure and properties.In this paper,based on machine learning,the supervised or unsupervised learning algorithms are purposefully designed to investigate the internal microstructure of material systems,system parameter-property prediction under small sample data,and the influence of material molecular structure on properties,etc.The relationship between structure and properties of materials is studied from multiple angles and levels,so as to accelerate the design and development of materials.The main research contents of the thesis are:(1)Cluster Envelope Shaping Method(CESM)is proposed to characterize the material microstructure and then analyze the properties.CESM consists of a distance-based clustering algorithm and a shape approximation algorithm to qualitatively and quantitatively characterize the microstructure of the system under different interaction strengths for the uniaxial stretching process in molecular dynamics simulations.When the intra-system interaction is weak,the clustering algorithm is used to observe the changes in the distribution of the system beads during the stretching process,and the changes in the microstructure under different system parameters are measured from the changes in the number and size of clusters.When the interaction within the system is strong,a shape approximation algorithm is used to achieve shape approximation based on the results of clustering,and the surface area and eccentricity of the approximated ellipsoid quantify the deformation of the material by the stretching action.(2)The mutual migration Gaussian process regression algorithm is proposed to achieve a multi-property prediction study of polymers based on small sample data.When constructing models for predicting the properties of different material system parameters,there is little data for fitting the model hyperparameters due to the long simulation time or the complexity of the material system experiments.The excellent regression prediction performance of Gaussian process in small sample tasks and the effect between learning properties using transfer learning to achieve data enhancement improve the prediction accuracy of Gaussian process regression models,provide guidance for complex material system parameter settings in molecular simulations,and reduce the time consumption of repeated simulations.In addition,the interactions of properties can be studied from the models constructed from the relationships between multiple properties,which helps to conduct in-depth analysis of the properties of complex polymer systems.(3)Based on graph networks,graph messgage passing networks and graph explaination algorithms are constructed to achieve property classification prediction and material molecular structure-property relationship interpretation for organic semiconductor materials.The graph information transfer network constructed by using the information transfer layer and gated loop unit extracts the feature information important to the property from the graph data composed of the molecular structure of the material and realizes the classification and prediction of the property.Based on the extracted feature information,the graph explaination algorithm weights the molecular structures and investigates the molecular substructures that have an important impact on performance.More,due to the diversity of molecular structures and the complexity of structureproperty relationships,algorithms such as clustering are used as an aid to discuss the results of structure-property relationships in a categorical manner.(4)An improved graph neural network algorithm based on molecular partitioning is proposed for better property prediction and more reasonable explanation of the structure-property relationship.Based on the composition mechanism of material molecules and the division of functional groups,an unsupervised learning method is constructed to segment the structure of the graph data composed of molecules.Combining the molecular segmented structure,a new graph neural network is designed to focus more on the local effects of functional groups,and the improved graph neural network has better prediction performance through experiments in databases such as solubility.At the same time,the combination of molecular segmentation and graph explaination algorithm leads the graph explainationg algorithm to find the substructure containing complete functional groups,which is more practical for the structure-property explaination and more consistent with the mechanism of molecular composition,and more meaningful for property analysis and new material design.