Low-Cost Wafer Inspection Using SEM-to-OM Image Simulation

Abstract

SiC Wafer inspection is crucial in semiconductor manufacturing and imaging is the one of the most popular methods of defect detection. Although scanning electron microscopy (SEM) provides high-resolution visualization of surface defects, its high cost of operation and low speed limit restrict largescale deployment. In contrast, optical microscopy (OM) offers a cost-effective and high-throughput alternative but suffers from limited availability of annotated OM datasets. In this paper, we propose a physics-inspired transformation module to generate synthetic OM-like images from SEM images. The proposed Transformation module incorporate key optical imaging characteristics including, resolution degradation, illumination variation, contrast adjustment, and noise injection to simulate realistic OM-like images. A simulated OM dataset is constructed from SEM images containing particle and surface patterns, which is subsequently used to train a convolutional neural network for defect classification. Experimental r esults s how t hat models trained on simulated OM data exhibit improved robustness on OM simulated images compared to those trained directly on SEM images. This paper shows the feasibility of the SEM-toOM simulation approach to bridge the data gap and facilitate low-cost wafer inspection solutions.