The ever-shrinking feature size of integrated circuits aggravates the subwavelength lithography gap, causing unwanted shape deformations of printed layout patterns. Although various resolution enhancement techniques (RETs) used to improve wafer printability are used to improve the imaging fidelity, certain layout regions may still be susceptible to the lithography process with pinching and bridging hotspots that may produce open or short circuits. Therefore, the identification of lithography hotspots is particularly important in physical verification. In this study, we propose a hotspot detection method to improve the precision and recall of pinching- and bridging-type areas by embedding squeeze-and-excitation networks (SENets) into a pretrained YOLOv3 model. We also address hotspot and non-hotspot data imbalances by data augmentation from a lithographic perspective. Experimental results on the 2012 International Conference on Computer-Aided Design (ICCAD 2012) dataset verify the merits of the proposed deep learning-based network.

YOLOv3 uses a single network to generate candidate regions within which the locations and classifications of objects are detected and identified. The training of YOLOv3 is more effective with a single network structure. SENet is an attention mechanism that focuses on the channel features. SENet provides information regarding the importance of each channel in the feature map, enabling the network to focus on important channels while suppressing less important channels. To better distinguish lithographic hotspots from non-hotspots, SENet was embedded in the YOLOv3 network architecture to improve the representation ability between different channels in the feature map. The structure of the improved YOLOv3 is shown in Fig. 4, where SENet is enclosed in the dotted box. Imbalanced datasets cause the network to focus more on learning the features of non-hotspots, thereby reducing the performance of hotspot detection. Considering the symmetry and light source in the lithography process, a change in the direction of the layout pattern does not alter its properties, and the number of layout patterns can be increased by flipping the original layout pattern. Fig. 5 shows the flipping data augmentation method.

In this study, the effectiveness of the lithographic hotspot detection task was verified by comparing the improved and the original YOLOv3 structures. In the experiments, the intersection over union threshold is set to 0.5, and the confidence threshold is set to 0.8. The experimental results are presented in Table 3. Although both the improved and original YOLOv3 networks show similar detection capabilities, the accuracy of the improved YOLOv3 is not significantly different from that of the original YOLOv3 on benchmarks 1,2, and 4, and the accuracy of the improved YOLOv3 is significantly greater than that of the original YOLOv3 on benchmarks 3, 5, and 6. To verify the effectiveness of the SENet proposed in this study in the lithographic hotspot detection task, a reference attention mechanism, convolutional block attention module (CBAM), was also tested by replacing the SENet in the dotted box of Fig. 4. CBAM is an attention mechanism that includes both spatial and channel attention mechanisms. The training and test settings of the CBAM detection network are consistent with those of the SENet detection network. The test results are listed in Table 4. The lithographic hotspot detection network embedded with the CBAM can accurately identify hotspots. In terms of accuracy, the performance of YOLOv3 embedded with the CBAM is the same as that of YOLOv3 embedded with the SENet on benchmarks 1, 3, and 5. For the other benchmarks, the accuracy of YOLOv3 embedded with the CBAM is significantly lower than that of YOLOv3 embedded with the SENet. According to this analysis, the proposed method, in which YOLOv3 is embedded with the SENet, outperforms the CBAM and the prevailing methods in the literature.

Lithographic hotspot detection is a key step in the physical verification process of very large-scale integration circuit (VLSI). The pattern on the wafer is easily affected by lithographic printing, and a sensitive layout pattern produces unwanted hotspots. The geometries of hotspots and non-hotspots are extremely similar, exacerbating overfitting in deep learning-based approaches. In this study, embedding the SENet in the YOLOv3 network can focus the network on important channels in hotspot and non-hotspot feature maps. By taking advantage of the symmetry of lithography imaging, the problem of data imbalance can be addressed by flipping the hotspot samples. In the ICCAD 2012 dataset, benchmark 1 was pretrained, whose training parameters were used as the initial weights of benchmarks 2 to 6 to accelerate the training speed and improve the performance of the network. The test results show that the average recall of the proposed method is 1.00, accuracy is 0.45, and F1 score is 0.62. Compared with the prevailing methods in the literature, the proposed deep learning network with the SENet improves the detection performance of lithographic hotspots.