Objective Lithographic tool is an important device for large-scale IC manufacturing. Its function is to transfer mask patterns into photoresists on wafers. Nowadays, the designed feature size of IC is below 10 nm, and the number of transistors of an IC is as high as tens of billions. With the demand for high integration and good performance, the physical design of IC continues to shrink, and lithographic printability has become one of the critical issues in IC design and manufacturing. Affected by the layout design and lithography process, the lithography results of some patterns in the layout are quite different from that of the target patterns, which results in short-circuit or open-circuit problems. These problems will cause lithography hotspots. In order to reduce lithography hotspots, hotspot detection and layout correction are carried out in turn in the layout design phase. The performance of the hotspot detection affects the period and yield of IC manufacturing. Hotspot detection is one of the important techniques for IC design and manufacturing. For available hotspot detection methods, the hotspot detection method based on lithography simulation is time-consuming, and the hotspot detection method based on pattern matching is invalid for unknown hotspot patterns. The hotspot detection method based on machine learning has good performance in speed and accuracy and has been widely studied. Transfer learning has been applied in the hotspot detection method based on machine learning and achieved positive model performance. Model performance and model training time affect the application of the hotspot detection method based on transfer learning. In this study, a lithography hotspot detection method based on a pre-trained VGG11 model is proposed. The proposed method helps to improve the model performance and model training time. Methods In this study, we adopt a transfer learning strategy for model training of hotspot detection. First, the ImageNet dataset is used to pre-train the VGG11 model, and the pre- trained VGG11 model is used as the model to be trained for hotspot detection. Then, the network architecture of the pre-trained VGG11 model is fine-tuned to make it suitable for hotspot detection. In the data preparation phase, pattern down-sampling and data balance are employed to prepare data for model training. In the model training phase, the strategy of preserving pre-trained model weights and freezing convolutional layers is adopted for model training. The trained model is suitable for hotspot detection. Results and Discussions In this study, ICCAD 2012 benchmark suite is used for model training and model tests. The ICCAD 2012 benchmark suite contains five subsets. According to the data-balance method, non-hotspot data are randomly sampled in the training data of the ICCAD 2012 benchmark suite. The randomly sampled non-hotspot data and all hotspot data constitute a complete training dataset, and the number of training data in the training dataset is 5054. The model test data are the five test datasets in the ICCAD 2012 benchmark suite. With the pre- trained VGG13, VGG16, and VGG19 models as references, VGG model comparison experiments are conducted. As shown in Fig. 5 and Table 2, the comparison results indicate that the comprehensive performance of the model obtained by the proposed method is better, and the average accuracy, recall, precision, and F-1 score of the model reach 98. 9%, 98. 2%, 89. 5%, and 93. 3%, respectively. In addition, the hotspot detection method based on the pre-trained VGG11 model requires a shorter model training time of only 279 s. The improvement of the model performance and model training time helps to improve the efficiency of hotspot detection. Compared with the hotspot detection method based on the pre-trained GoogLeNet model and the pre-trained VGG16 model, the results show that the proposed method has better comprehensive model performance. As shown in Fig. 6 and Table 4, on the premise that the hotspot detection has a considerable recall, the proposed method can significantly improve the precision and F-1 score. Tests are carried out to release convolutional layers with different numbers for model training, and the results show that the release of the convolution layers has a slight effect on the model performance of the proposed method. Conclusions A lithography hotspot detection method based on a pre-trained VGG11 model is proposed in this study. A transfer learning strategy is adopted for model training. The proposed method uses a VGG11 network trained by the ImageNet dataset as the pre-trained model, and the network architecture of the pre-trained VGG11 model is fine-tuned to make it suitable for hotspot detection. Model training is performed by using a strategy of preserving pre- trained model weights and freezing convolutional layers. ICCAD 2012 benchmark suite is used for model training and model tests. Compared with that of available methods, the results show that the model of the proposed method has better comprehensive performance and requires less model training time. The average accuracy, recall, precision, and F-1 score of the model reach 98. 9%, 98. 2%, 89. 5%, and 93. 3%. In addition, the model training time of the proposed method is only 279 s. The proposed method helps to improve the efficiency of hotspot detection and shortens the period of integrated circuit manufacturing.
