Semiconductor image analysis using data-efficient machine learning

Recent advancements in AI have intensified interest in automating Transmission/Scanning Electron Microscopy (TEM/SEM) image analysis for semiconductor metrology and inspection (MI). However, the scarcity of annotated datasets in the industry poses a critical challenge, as model performance heavily relies on the quantity of precisely labeled training data. To address this limitation, we propose a data-efficient machine learning framework optimized for structural and categorical semiconductor image analysis. Our approach integrates three key innovations: (1) Transfer learning of a foundation model for semiconductor image segmentation, which outperforms generic counterparts by 0.1920 in mIoU; (2) Semi-supervised learning with TEM-specialized augmentation strategy with limited labeled data, achieving competitive mIoU of 0.877 even with a few labeled examples; and (3) Self-supervised contrastive pre-training for SEM defect image classification, leveraging magnification-invariant consistency to enhance feature robustness. Evaluated under limited-data regimes, our framework demonstrates superior performance over conventional methods while significantly reducing annotation dependency. This solution directly tackles the semiconductor industry’s persistent challenge of high labeling costs and rapidly expanding unlabeled data pools, offering a practical pathway for AI adoption in semiconductor developments.