Generative Machine Learning

This work demonstrates that GAN-based models, especially StyleGAN2-ADA, can generate statistically representative synthetic microstructures even with limited datasets, addressing a major bottleneck in materials data scarcity. Unlike prior studies relying on large datasets or qualitative validation, this paper emphasizes rigorous quantitative validation, using metrics such as statistical, (FID, KID) morphometric descriptors (phase fraction, lamellar features), and two-point correlations. The key takeaway is that generative models can faithfully reproduce the statistical distribution of real microstructures, making them viable tools for data augmentation, microstructure design, and accelerating Process-Structure-Property modeling workflows. 

This paper introduces the concept of microstructural fingerprinting using variational autoencoders (VAEs) as a way to overcome the limitations of traditional, hand-crafted descriptors (e.g., grain size, phase fraction) in capturing complex microstructures. By encoding Ti-6Al-4V micrographs into a continuous latent space, the authors demonstrate that these learned representations are smooth, interpretable, and correlated with key physical features such as volume fraction and grain size. Crucially, the ability to reconstruct and interpolate microstructures from this latent space provides both interpretability and generative capability, establishing VAEs as a powerful tool for building quantitative process–structure–property (PSP) linkages and enabling data-driven materials design.