Research On Material Grain Representation Method Based On EBSD Knowledge Graph Representation Learning

Polycrystalline materials are widely found in industry and daily life.Common metals such as iron alloys,copper alloys,magnesium alloys and aluminum alloys are generally polycrystalline materials.Polycrystalline materials are widely used in life,construction,automobiles,aerospace,medicine and other fields,occupying an indispensable position in human society.Therefore,finding polycrystalline materials with exceptional properties has always been an important topic for materials researchers.The material properties are primarily determined by their microstructure genes,and material research is also a process of exploring the quantitative relationship between tissue structure and properties.At present,the structure-property research of polycrystalline materials mainly includes: artificial qualitative analysis and physical modeling.The method based on qualitative analysis is to observe specific features through images or statistical functions,etc.,and then establish qualitative structure-property relationships,such as texture analysis,grain size analysis;another method is to simulate and calculate the properties of materials by establishing physical models,such as viscoplastic self-consistent polycrystalline plastic model,crystal plastic finite element method.Manual analysis is time-consuming and labor-intensive and requires extensive expert experience.However,methods based on traditional physical modeling are not only complex to model but also have extremely low computational efficiency on large-scale grain structures.With the advent of data science,the application of machine learning and data mining technology in material computing has developed rapidly.At present,for polycrystalline materials,there is still a lack of a more general data model computing method.The main reason is the absence of an effective and general structural representation description.Aiming at the above problems,this thesis proposes a material grain representation method based on EBSD knowledge graph representation learning,and applies this method to the structure-property prediction of materials.The main work includes the following:(1)This thesis designs a digital representation of attribute grain knowledge graph based on EBSD.The representation is based on the knowledge graph technology,describing the grains and grain boundaries in the polycrystalline structure through edges and nodes,and constructing a heterogeneous graph to embed different grain interaction relationships and heterogeneous attribute associations in the polycrystalline structure,which can effectively restores structural information describing polycrystalline materials and supports digital storage as well as machine learning.(2)This thesis proposes a representation learning model based on the EBSD grain knowledge graph,which captures the rich heterogeneous information in the grain knowledge graph and enhances the expression of key property features in polycrystalline structures by using double-layer graph attention network and attribute message passing learning.This model enables a graph embedding of the polycrystalline structure to be obtained for downstream structure-property prediction tasks.(3)This thesis designs a data-driven material structure-property calculation method based on the polycrystalline structure feature embedding obtained by the EBSD grain knowledge graph representation learning.The computational method combines digital representation and representation learning method to complete the establishment of polycrystalline materials from characterization to structure-property gene relationship,which provides a novel idea and method for materials informatics research.In this thesis,the proposed method is experimentally tested on magnesium alloy data.The experimental results show that this method is superior to the machine learning method based on statistical descriptors,which proves its effectiveness and superiority.At the same time,we apply the method to other polycrystalline material system,and also achieved good characterization results and predictive performance,further demonstrating the effectiveness and universality of the method proposed in this thesis.