SiliconBASE: multi-task baseline model for semiconductor metrology and inspection applications

Metrology and Inspection (MI) play a crucial role in semiconductor manufacturing. Recently, existing MI methods have faced significant challenges due to increasing noise and reduced clarity in microscopic images. While Convolutional Neural Networks (CNNs) have shown promising results, their adoption in MI remains limited due to the need for extensive labeled datasets and the absence of widely-used backbone networks optimized for semiconductor images. To address these limitations, we propose SiliconBASE, a novel unsupervised learning architecture optimized for MI tasks. Our approach enables robust feature extraction by utilizing unlabeled data, providing a backbone model adaptable to various downstream applications, including object detection and segmentation. Furthermore, we introduce reference object detection (ROD) and reference object segmentation (ROS), which allow flexible object recognition without predefined categories, significantly improving generalization ability. Extensive experiments show that SiliconBASE achieves superior performance compared to supervised models such as Mask-RCNN trained on limited datasets, even surpassing human annotation in challenging cases. SiliconBASE establishes a scalable, high-performance baseline for MI applications in the semiconductor industry. In future work, we aim to expand its applicability to larger datasets and additional downstream tasks, further enhancing its generalization and efficiency.