A Novel Optical Proximity Correction Machine Learning Model Using a Single-Flow Convolutional Feedback Networks With Customized Attention

In semiconductor fabrication, any deviation leads to significant mistakes in the result. Thus, the proximity effect is a critical issue that must be solved. In the past, optical proximity correction was constructed by fabrication experience and physics formula models, resulting in difficulties when the technology node shrinks. As a result, optical proximity correction with machine learning models is highly expected to solve the issue in recent years. Due to the unique feature in optical proximity correction, single-flow convolutional feedback networks with customized attention layer are proposed to compete with widely used U-Net or U-Net with attention layer, which is the current mainstream in image-to-image machine learning tasks. The customized attention layer is used to replace the conventional attention layer. The proposed model with a customized attention layer has improved metrics compared to U-Net or U-Net with an attention layer. Compared the proposed model to U-Net with a cross-attention layer, we observe 3.74% improvement of modified mean pixel accuracy in the two-bar dataset, 0.9% improvement of modified mean pixel accuracy in the tri-bar dataset, 3.76% improvement of modified mean pixel accuracy in the polygon dataset and 2.06% improvement of modified mean pixel accuracy in the GAN400 dataset.