Fusion materials in magnetic fusion devices must tolerate extreme loading conditions. Tungsten remains the leading plasma-facing component (PFC) material, yet the microstructural variability governing its degradation under harsh conditions, including crack initiation, porosity evolution, and thermal fatigue, remains costly to characterize at scale. We developed a data-efficient generative workflow to synthesize realistic scanning electron microscopy (SEM) microstructures of tungsten for physics-aware data augmentation and rapid hypothesis testing under fusion-relevant environmental conditions. Starting from 3200 SEMs acquired from electron-beam (e-beam) heat flux exposure studies on tungsten, we tiled each image into fixed-field 256 × 256 grayscale patches and trained two models: a baseline model, conditional-GAN (c-GAN) and a style-based model, conditional StyleGAN2 with Adaptive Discriminator Augmentation (c-StyleGAN2-ADA). The latter adapts augmentation strength during training and is well suited to the small-data regime. Fidelity was assessed with a physics-aware validation suite: (i) image distributional similarity to real SEMs via Fréchet Inception Distance (FID) and Kernel Inception Distance (KID); (ii) microstructure realism via grain-size statistics derived from classical image analysis and comparison of grain-area statistics via Kolmogorov-Smirnov (KS) and Earth Mover’s Distance (EMD); and (iii) anti-memorization screening using nearest-neighbor searches with Learned Perceptual Image Patch Similarity (LPIPS). Notably, our trained c-StyleGAN2-ADA generator reproduced grain-size distributions that closely followed the real data while maintaining diversity and avoiding trivial copies, outperforming the c-GAN on both perceptual and physics-aware metrics. The approach yields physically plausible microstructures on demand and provides a basis to seed multi-scale degradation models, uncertainty analyses, and “virtual experiments” for PFC design when direct measurements are scarce.
