Process-conditioned transmission electron microscopy image generation from optical critical dimension metrology

초록

The purpose of this study is to examine whether non-destructive in-line optical critical dimension (OCD) can support transmission electron microscopy (TEM) structural and metrological tasks such as morphological screening and approximate critical dimension assessment by synthesizing TEM images conditioned on OCD measurements. In semiconductor manufacturing, nanometer-scale metrology is central to yield and time-toramp. However, the high cost, long turnaround time, and the destructive nature of TEM limit measurement frequency and coverage. To this end, we develop OCD2TEM, a deep learning-based conditional generative framework that bridges the dimensional gap between low-dimensional OCD data and high-dimensional TEM imaging. The key component is an OCD encoding module that uses attention mechanisms to transform sparse optical measurements into rich conditioning signals. Using an 8-nm OCD-TEM paired dataset, we synthesize TEM images from OCD to assess how closely they reproduce real source-drain morphology and critical dimensions. Across quantitative image-quality evaluations and critical dimension measurements, the synthesized images preserve source-drain structural morphology on our 8-nm dataset and tasks. Moreover, performance does not degrade substantially as training data are reduced, indicating robustness under data scarcity and favorable data efficiency. Our findings indicate a practical, non-destructive route to TEM-based structural monitoring based on realistic, scarce paired data. This approach supports broader wafer-level screening and process control.

키워드