Litho-NeuralODE 2.0: Improving hotspot detection accuracy with advanced data augmentation, DCT-based features, and neural ordinary differential equations

Abstract

It has proved that the application of deep neural networks has advantage in lithographic hotspot detection, which is vital in the physical verification flow to reduce manufacturing yield loss. In this paper, we employ the discrete cosine transform (DCT)-based feature extraction method along with parameter search to compress the layout image to achieve higher classification accuracy and speed up the training process. To further improve the classification performance, data augmentation techniques addressing the imbalanced dataset problem along with neural ordinary differential equations based Litho-NeuralODE 2.0 framework with improved loss function have utilized in the work. Experimental results demonstrate that the proposed framework outperforms the state-of-the-art works with the lowest misses of 7 and the highest accuracy of 98.9% on average.