A physics-informed synthetic-to-experimental framework for few-shot structural segmentation of HRTEM images

High-resolution transmission electron microscopy (HRTEM) can reveal nanoscale structural heterogeneity, yet converting weak and ambiguous image contrast into reproducible multiclass structural maps remains difficult because reliable pixel-level annotations are scarce. Existing HRTEM segmentation workflows therefore often remain limited to binary foreground-background separation or other simplified tasks, rather than simultaneous parsing of crystalline regions, grain boundaries, amorphous or structurally unclear regions, and background. Here we report a physics-informed synthetic-to-experimental framework that enables few-shot multiclass HRTEM segmentation. The key advance is a Domain Construction-Relaxation-Imaging (DCRI) framework that generates synthetic HRTEM images with strictly co-registered structural labels, providing scalable supervision for synthetic pretraining. In this two-stage workflow, DCRI pretraining provides transferable structural priors, while few-shot experimental fine-tuning adapts these priors to real HRTEM images. Using Au nanoparticles as the primary model system, we show that this strategy produces coherent four-class segmentation under low-label conditions, whereas experimental-only training remains unstable. The benefit is retained across representative U-Net-family backbones and after few-shot adaptation to Au STEM images, shows partial transfer to annotated PbS HRTEM images, and further supports SAM-based foundation-model adaptation. These results establish a physics-informed synthetic-to-experimental framework as a practical route to alleviate the annotation bottleneck in few-shot structural segmentation of HRTEM images.