Diffusion Model Generates Realistic TEM Images from Only 15 Samples

Researchers Johannes Boehm and Bappaditya Dey have published a method using Denoising Diffusion Probabilistic Models (DDPMs) to generate synthetic Transmission Electron Microscopy (TEM) images for semiconductor metrology. Their key innovation is a progressive patch-based training strategy that starts with low-resolution patches and scales to full images, enabling training from scratch with only 15 real samples—a dramatic reduction from typical data requirements. The model incorporates custom TrivialAugment data augmentation, cross-process domain transfer, classifier guidance, and RePaint-style inpainting to preserve the specific noise patterns, structural details, and stochastic variability that real TEM images exhibit.

The results are striking: synthetic images achieve Multi-Scale Structural Similarity Index (MS-SSIM) greater than 0.98, meaning they are nearly indistinguishable from real TEM images by both algorithmic metrics and expert assessment. Beyond image generation, the authors repurpose the DDPM feature representations for segmentation tasks, producing coherent region masks that can directly train defect detection and metrology models. This approach addresses a critical bottleneck in advanced semiconductor manufacturing, where destructive sample preparation, slow imaging, and high costs severely limit available datasets. By generating physically and statistically realistic TEM images from minimal real data, the method could accelerate ML-driven process control and quality assurance in chip fabrication.