SinSEMI: A One-Shot Image Generation Model and Data-Efficient Evaluation Framework for Semiconductor Inspection Equipment

TL;DR

SinSEMI addresses data scarcity in semiconductor optical image generation by proposing a one-shot, multi-scale flow-based model guided by LPIPS energy. It demonstrates that training-free LPIPS guidance during sampling improves perceptual fidelity and diversity while remaining computationally efficient. The authors also introduce a data-efficient evaluation framework requiring only two reference images to assess visual quality, quantitative similarity (SIFID/LPIPS), and downstream defect segmentation performance. Across experiments on line-pair structures, SinSEMI achieves strong fidelity and meaningful variation, enabling effective AI model training with minimal real data.

Abstract

In the early stages of semiconductor equipment development, obtaining large quantities of raw optical images poses a significant challenge. This data scarcity hinder the advancement of AI-powered solutions in semiconductor manufacturing. To address this challenge, we introduce SinSEMI, a novel one-shot learning approach that generates diverse and highly realistic images from single optical image. SinSEMI employs a multi-scale flow-based model enhanced with LPIPS (Learned Perceptual Image Patch Similarity) energy guidance during sampling, ensuring both perceptual realism and output variety. We also introduce a comprehensive evaluation framework tailored for this application, which enables a thorough assessment using just two reference images. Through the evaluation against multiple one-shot generation techniques, we demonstrate SinSEMI's superior performance in visual quality, quantitative measures, and downstream tasks. Our experimental results demonstrate that SinSEMI-generated images achieve both high fidelity and meaningful diversity, making them suitable as training data for semiconductor AI applications.

Figures (5)

• Figure 1: Overview of SinSEMI's forward and backward processes: The forward process (left) progressively diffuses the optical image from the finest scale (Scale=N-1) to the coarsest scale (Scale=0) while building a pyramid of downsampled images, whereas the backward process (right) generates images in a coarse-to-fine manner through iterative denoising starting from Gaussian noise.
• Figure 2: Optical simulation images of line pairs. The left image shows a line pair with a bridge defect, while the right image shows a line pair without defects
• Figure 3: Visual comparison of generated samples across different generative models for line pairs. The training image for each structure is shown on the left, followed by samples generated by different models.
• Figure 4: Spatial distribution of defects generated by different generative models for line pairs. Heat maps show accumulated defect locations from 1000 generated samples per model, where darker colors indicate higher frequency of defects at each location.
• Figure 5: Defect segmentation results for SinSEMI+Energy. (Left) Testing optical image. (Middle) Ground truth defect segmentation. (Right) Predicted defect segmentation by the SinSEMI+Energy model.