Physics-aware losses keep grain boundaries real when AI generates alloy microstructures

bravo_abad · x · 2026-09-30

When generating microstructures for high-strength, ductile nickel-based alloys, Liao et al. note that blurring a single grain boundary changes few pixels but can hide a consequential physical error. Their pipeline: a variational autoencoder generates initial microstructure images, then a diffusion model refines them. The borrowable trick is the training objective: alongside pixel-wise reconstruction error, they add an edge loss emphasizing grain boundaries and a structural similarity loss preserving the spatial organization of crystal orientations. With pixel-wise error alone, reconstructions blur boundaries and distort the physics.

Original post →

More from Research

Research channel →