Semi-supervised lithography hotspot detection based on feature fusion and residual attention

Given that traditional lithography hotspot detection methods based on semi-supervised learning struggle to meet the detection accuracy requirements of advanced integrated circuit (IC) manufacturing. To address the above challenges, a semi-supervised detection method based on feature fusion and residual attention is proposed in this paper. The method employs two inception modules as the multi-scale feature fusion module (MFF). These modules work in parallel to combine features from different layout scales. A residual attention module (RA) based on the convolutional block attention module (CBAM) is introduced, and a new neck network called RANeck is constructed using the RA module. The model utilizes the original layout features by constructing a joint multi-task network for classification and clustering through RANeck. The introduction of CBAM allows the model to focus more on important feature channels, thereby achieving more precise information filtering during feature processing. Additionally, a weighted cross-entropy loss function dynamically adjusts weights during the training process based on the number of lithography hotspots and nonhotspots, mitigating data imbalance effects and reducing false alarms. This method effectively leverages a large number of unlabeled data for training, improving the accuracy of lithography hotspot detection in the case of insufficient labeled data. The experimental results show that compared with the existing semi-supervised lithography hotspot detection methods, the proposed method has improved accuracy, false alarm, F1 score, and overall detection simulation time using 10 %–50 % of training data on the ICCAD 2012 contest benchmarks by 3.48 %, 22.03 %,12.76 %, and 20.26 %, respectively.