Lithography hotspot detection method based on convolution kernel reconstruction

BackgroundIn advanced technology nodes of large-scale integration circuits, lithography hotspot detection not only significantly impacts manufacturing economics but also critically determines long-term device reliability and circuit functional integrity. Conventional methods relying on rule-based approaches and simulation verification exhibit inherent limitations in operational efficiency and the capability to capture complex nonlinear features.AimTo overcome missed detections and false positives of lithography hotspots caused by deep semantic feature loss in fixed convolution kernel, this work simultaneously resolves severe classification bias from cross-entropy loss under class imbalance, preventing significant degradation in minority-class classification performance.ApproachWe introduce a residual network architecture based on Kolmogorov-Arnold Networks (KAN). First, conventional convolution kernels are replaced with KAN modules to enhance high-order nonlinear mapping capabilities through universal approximation properties, whereas tensor decomposition strengthens cross-channel feature interactions. Second, the reconstructed residual blocks integrate adaptively generated KAN features with original skip-connections via element-wise summation, alleviating gradient attenuation during deep network training. Finally, a weighted binary cross-entropy loss function dynamically adjusts weights according to sample distribution, effectively mitigating model bias.ResultsThe proposed algorithm was evaluated on the ICCAD 2012 benchmark dataset, achieving a recall of 96.65%, a precision of 98.55%, and an F1-score of 97.59%.ConclusionsCompared with other deep learning-based methods, the proposed approach exhibits enhanced representational capacity, achieving superior precision and F1-score.