High-Fidelity Synthetic Transmission Electron Microscopy Image Generation Using Diffusion Probabilistic Models for Data-Limited Semiconductor Metrology

This study addresses the critical data scarcity in semiconductor metrology by leveraging Denoising Diffusion Probabilistic Models (DDPMs) to generate synthetic Transmission Electron Microscopy (TEM) images. Traditional physics-based simulations often struggle with the complexity of real-world device structures, while standard generative models frequently fail to capture the specific noise profiles and stochastic variability inherent in TEM imaging. The authors propose a progressive, patch-based training strategy that scales from low-resolution patches to full-field images. This approach allows the model to learn fine-grained textures at smaller scales before integrating global structural relationships at higher resolutions, enabling effective training from extremely limited datasets (as few as 15 images).

The framework integrates several advanced techniques to ensure physical and statistical realism. A custom adaptation of the TrivialAugment (TA) algorithm, tailored for grayscale TEM data, provides robust data augmentation without compromising structural integrity. The model also employs v-prediction parameterization and a min-SNR-gamma loss weighting scheme to improve training stability and sample quality. Beyond image generation, the authors demonstrate that the intermediate feature representations of the DDPM can be repurposed for unsupervised segmentation, providing coherent region masks that serve as efficient starting points for downstream metrology tasks.

As semiconductor nodes shrink to 2 nm and beyond, the demand for high-quality TEM data for defect inspection and process control has surged. However, the destructive nature of sample preparation and the high cost of imaging create a significant bottleneck for machine learning development. By providing a method to generate high-fidelity, domain-consistent synthetic data, this research enables the training of more robust downstream ML models for defect detection and metrology, ultimately reducing the reliance on scarce and expensive experimental data.