Self-Supervised Siamese Transformer for Surface Defect Segmentation in Diamond-Wire-Sawn Mono-Crystalline Silicon Wafers

Abstract

Silicon wafers are the most common semiconductor substrate, and their quality plays a decisive role in the service performance of semiconductor devices. Diamond wire sawing is the first and key machining process for manufacturing silicon wafers. However, the process faces a challenge in achieving high efficiency and high quality due to the brittle characteristics of silicon wafers, which unavoidably generate surface defects like scratches and pits during the wire sawing process. In this article, a novel self-supervised Siamese transformer (S3Transformer) model is proposed to segment silicon wafer surface defects which includes two modules, namely Siamese representation learning (SRL) and multiscale representation learning (MRL). MRL is responsible for learning different scales of features and SRL serves as an auxiliary network to improve the small-scale feature learning by adding a Siamese loss. In addition, a novel CNN-based projector in the SRL module is designed for computation cost-saving and deeper feature learning. The surface segmentation dataset of silicon wafers is established to validate the model. The results show that the S3Transformer yields better silicon wafer surface defect segmentation performance compared to other state-of-the-art models.