SO-Diffusion: Diffusion-Based Depth Estimation From SEM Images and OCD Spectra

As the semiconductor industry evolves, semiconductors are becoming smaller and more complex. Therefore, reconstructing the 3-D structure of the semiconductor is very crucial. To predict the structure of the semiconductor, there are some studies to estimate depth from a scanning electron microscope (SEM) image. However, it is difficult to estimate the depth of the semiconductor only from a single SEM image because the top-view SEM image has insufficient information of depth. Thus, we propose SO-Diffusion, a novel framework to estimate the depth from the SEM image as using optical critical dimension (OCD), which provides more 3-D information. SO-Diffusion is based on a diffusion network to extract features from both SEM images and OCD spectra, fusing these distinct features to accurately estimate the depth. Specifically, we introduce data encoders for SEM image and OCD spectrum, and especially, we design the novel spectrum encoder, SEFO, for accurate feature extraction of OCD spectra. We conducted a lot of experiments for diverse environments and demonstrated new state-of-the-art results, with performance gains of up to 56% or more. Our work presents a novel method for structural metrology of the semiconductor, leading to a world of finer grained semiconductors.