Geometrical shape learning as basis for compact microstructure representations and microstructure-properties linkages

Process-structure-properties linkages play a major role in materials and process engineering. Nowadays, such linkages are often established on the basis of experimental data and simulation data using machine learning approaches. For this purpose, typically, state-of-theart feature extraction methods, such as principal component analysis, are used in combination with regression models, such as neural networks. For complex spatially-resolved microstructure representations convolutional neural networks are often used, which are, however, very data-intensive and not explainable. In this work, we present a novel approach based on geometrical shape features that allows for compact microstructure representations, even with a small amount of data, and is, furthermore, explainable. In addition, the presented approach maps the identified features to a latent feature space that is not dependent on the data.