Integrating CNN and Transformer for efficient lithography hotspot detection

As the complexity of the lithography process in the semiconductor industry increases, so do the patterns that can lead to manufacturing defects. This has made the prediction and correction of lithography hotspots during the design phase a crucial task. However, traditional hotspot detection methods often struggle with detecting novel patterns or require significant computational resources. In this paper, we propose a hybrid approach that combines Convolutional Neural Network (CNN) with Vision Transformer (ViT) through heterogeneous knowledge distillation. This method transfers knowledge from a CNN-based teacher to a ViT-based student, combining the strengths of both architectures. CNN is known for its proficiency in extracting local spatial features, while ViT excels at capturing global dependencies across images. Our proposed framework was evaluated using the ICCAD 2012 CAD contest benchmark. It outperforms standalone trained CNN and ViT models, showing up to a maximum of 32.8% improvement in F1-score. Compared to the state-of-the-art CNN models, it achieves a 9.8% improvement in F1-score and a remarkable 93.3% reduction in false alarms. These results demonstrate the potential of Transformer models for lithography hotspot detection, offering an efficient and scalable solution that could improve manufacturing yield when applied to semiconductor design verification processes.