Re-parameterizable deep layer aggregation network for lithography hotspot detection

BackgroundWith the continuous scaling down of integrated circuit feature sizes, the mismatch between the resolution of lithography systems and the feature dimensions leads to undesired shape distortions in wafer imaging, which can cause circuit malfunctions. These distorted regions are referred to as lithography hotspots (LHSs). Therefore, hotspot detection is crucial for ensuring the manufacturability and yield of integrated circuits.AimWe aim to develop a high-performance hotspot detection framework that enhances feature representation, reduces false alarms, and accelerates inference. It addresses the limitations of existing methods in detecting complex and fine-grained hotspot patterns under advanced process nodes.ApproachWe propose a re-parameterizable deep layer aggregation network that employs deep layer aggregation (DLA) to iteratively and hierarchically fuse multi-level features for improved hotspot recognition. The backbone integrates re-parameterizable convolutional blocks to speed up inference, whereas a multi-scale triple fusion (MSTF) module is introduced to further strengthen multi-scale feature integration.ResultsWe evaluated our method on standard LHS benchmarks. It outperforms state-of-the-art models in accuracy, false alarm reduction, and F1 score while cutting end-to-end runtime. Compared with prior approaches, our method shows measurable gains across all key metrics. Specifically, it delivers notable improvements in both detection performance and efficiency.ConclusionsThe proposed DLA-based re-parameterizable architecture offers an effective and efficient approach for detecting LHSs, improving both accuracy and computational efficiency. We point to a practical direction for design that considers manufacturability and could be extended to future studies on hotspot prediction in emerging technology nodes and more complex patterning processes.