Lithography hotspot detection through multi-scale feature fusion utilizing feature pyramid network and dense block

Abstract

Lithography hotspot (LHS) detection is crucial for achieving manufacturability design in integrated circuits (ICs) and ensuring the final yield of ICs chips. Recognizing the challenges posed by conventional deep learning-based methods for lithographic hotspot detection in meeting the demands of advanced IC manufacturing accuracy, this study introduces an LHS detection approach. This approach leverages multiscale feature fusion to identify defects in lithographic layout hotspots accurately. This method incorporates the convolutional block attention module into the backbone network to enhance the focus of the model on the layout area. Additionally, a feature pyramid is employed to merge deep and shallow features from the layout pattern, significantly enhancing the capability of hotspot detection network to extract both image and semantic features. Concurrently, by utilizing a dense block that directly interconnects various layers, the network gains the capacity to capture the correlation between low-level and high-level features, thereby enhancing the perceptual capabilities of the model. Experimental results demonstrate the superiority of the algorithm across accuracy, false alarm, F1 score, and overall detection simulation time compared to alternative lithographic hotspot detection algorithms.