Structure-property modeling scheme based on optimized microstructural information by two-point statistics and principal component analysis

Construction of the structure-property (SP) relationship is an important tenet during materials development. Optimizing microstructural information is a necessary and challenging task in understanding and improving this linkage. To solve the problem that the experimental microstructures with a small size usually fail to represent the entire sample structure, a data-driven scheme integrating two-point statistics, principal component analysis, and machine learning was developed to reasonably construct a representative volume element (RVE) set from the small microstructures and extract optimized structural information. Based on the elaborate quantitative metrics and method, this kind of RVE set was successfully constructed on an experimental microstructure dataset of ferrite heat-resistant steels. Moreover, to remove redundant information included in two-point statistics, the critical threshold of the tolerance factor related to the coherence length in microstructures was determined to be 0.005. An accurate SP linkage was finally established (mean absolute error < 6.28 MPa for yield strength). This scheme was further validated on two other simulated and experimental datasets, which proved that it can offer scientific nature, reliability, and universality compared to traditional strategies. This scheme has a bright application prospect in microstructure classification, property prediction, and alloy design.