AI-Powered Defect Detection using Deep Learning: A Pattern-Agnostic Faster R-CNN Approach for SEM Images with GPU Acceleration

Semiconductor manufacturing requires highly precise defect detection to ensure product quality and yield. This paper presents a deep learning-based defect detection framework using Faster R-CNN to identify and classify defects in Scanning Electron Microscope (SEM) images, specifically within the modeling flow, which is used to simulate and predict semiconductor process behavior based on metrology data. The process begins with manufacturing test patterns through the photolithography process, followed by metrology analysis using a SEM machine. If a defect is present in the SEM image, the associated metrology data may become unreliable, potentially degrading the accuracy of the simulation models built on this data. Using such erroneous data in model creation could lead to inaccurate predictions and weaker models.To address this issue, our method filters out defective data, ensuring that only high-quality, defect-free measurements contribute to model development. This leads to more reliable semiconductor process simulations and improved process decisions. Leveraging the proven performance of Faster R-CNN—particularly its effectiveness in detecting small, low-contrast anomalies—the model generalizes across diverse layouts without requiring pattern-specific training. It achieves a mAP of 83% at IoU = 0.5, significantly reducing manual inspection time and improving defect classification accuracy. Our approach enables scalable, automated inspection of SEM data, enhancing the overall efficiency, robustness, and precision of the semiconductor modeling and manufacturing pipeline.