From microstructural images to effective material properties
Feature extraction based on 2-PCF
While machine learning today relies heavily on convolutional neural networks when it comes to image analysis, the two-point correlation functions (2-PCF) can complement them efficiently: Analytical homogenization methods are often based on (partial) information on the two-point statistics which emphasizes its relevance for the effective microstructure -property relationship.
We account for the 2-PCF by studying a large variety of random microstructures and extracting the dominant information by means of an incremental snapshot proper orthogonal decomposition (POD)/principal component analysis (PCA). The resulting reduced parameters are used as input features for subsequent learning tasks alongside, e.g., the phase volume fractions of the material contrast.
Funding information and acknowledgement
Funded by Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC 2075 - 390740016. We acknowledge the support by the Stuttgart Center for Simulation Science (SC SimTech). |