Towards automatic feature extraction and sample generation of grain structure by variational autoencoder

Granularity is one of the most predominant structures for polycrystalline materials, from geological compounds to technical high-performance alloys. In particular, grain structures in metal alloys enable custom-tailored material properties if their formation can be modeled and simulated accurately to analyze, characterize, and generate granular materials. Despite significant advances in the field, established physics-based grain-growth models still need to improve to match real-life experimental data accurately. Instead of relying on physical knowledge, data-driven models provide a complementary way to study grain structures using only the information in the data. However, many machine learning methods require supervised learning on information that is challenging to measure or numerically expensive to compute. In this work, we propose a new strategy that uses unsupervised deep learning on synthetic physics-based data to generate realistic granular structures with customizable features and properties (grain size/number). A variational autoencoder learns to compress grain structures into a low-dimensional latent space, derive data-driven features that characterize grain structures, and use these features to generate new structures with representative grain size distributions in a computationally efficient way. These data-driven features can complement the physics-based features derived from classical materials research and help identify new characterizing properties. Thus, this work represents a prototype application of bridging machine learning methods to material characterization and generation.