An Improved YOLOv5 Model for Lithographic Hotspot Detection

The gap between the ever-shrinking feature size of integrated circuits and lithographic manufacturing ability is causing unwanted shape deformations of printed layout patterns. The deformation region with problematic imaging, known as a hotspot (HS), should be detected and corrected before mask manufacturing. In this paper, we propose a hotspot detection method to improve the precision and recall rate of the fatal pinching and bridging error due to the poor printability of certain layout patterns by embedding a spatial attention mechanism into the YOLOv5 model. Additionally, transfer learning and pre-trained techniques are used to expedite training convergence. Simulation results outperform the depth-based or representative machine learning-based methods on the ICCAD 2012 dataset with an average recall rate of 1, a precision rate of 0.8277 and an F1-score of 0.9057.