MFRLNet: A Layout Hotspot Detection Network Based on Multi-feature Representation Learning

Layout hotspot detection (LHD) aims to localize potential hotspot regions from complex layouts that may cause exposure distortions, which is crucial in the physical verification process. With the development of technology nodes to the physical limit, traditional methods can only retrieve hotspot patterns with simple structures and struggle to handle more complex patterns, thus failing to adequately address practical requirements. In this study, we model the layout hotspot detection task as a process of multi-feature representation learning, and accordingly design a novel layout hotspot detection network based on multi-feature representation learning (MFRLNet). Specifically, we design a reconstruction network (Rec Net) and an adaptive cross-layer feature fusion network (ACFF Net) to perform feature reconstruction and selectively fuse multi-scale reconstructed features in MFRLNet, respectively. Motivated by the multi-feature modeling and interaction (MFM-I), Rec Net extracts visual and geometric features from various layers using MFM-I, and conducts interactive learning to enhance the representational capacity of the network. Thereby, the module facilitates learning by the subsequent Classifier. Moreover, due to the extreme imbalance between hotspots and nonhotspots, we propose a two-stage training strategy to alleviate the issue of the network overfocusing on nonhotspot features. Experiments on the public datasets ICCAD 12 and ICCAD 19 demonstrate that our method achieves superior performance. Our code will be released at https://github.com/ChenHan-AnHui/MFRLNet.