Subpixel keypoint localization and angle prediction for lithography marks based on deep learning

The alignment precision in lithography determines the pattern transfer quality, whereas the initial angular deviation between the wafer and mask leads to the empty-window problem. Therefore, identifying the angular deviation from a single-view alignment mark image and reducing it below the threshold of empty-window is urgent. We propose a deep learning-based method to achieve high-accuracy angle estimation under low-resolution and blurry contours. The proposed network is a multitask network with a CNN as backbone and a parameter-free SVD-based head. It can simultaneously predict subpixel coordinates of alignment marks’ keypoints and registration angle. The network takes image patches rather than the entire original image as input to let itself only focus on the unobstructed keypoint regions. A multicomponent loss function based on angular relationships is also introduced. The training result verifies that the network can learn keypoint localization indirectly through angle prediction optimization. Application experiments demonstrate that the proposed method achieves an RMSE (root mean square error) of 0.093 and R $$^2$$ 2 (R-squared) of 0.976 in angle prediction error, effectively addressing the empty-window problem. Furthermore, compared to other angle estimation methods based on keypoint registration, line detection, template matching, and manual select points, our method demonstrates higher accuracy and better stability. The code and models have been publicly available at https://github.com/YuLungLee/HRNet8-SVD . Our model is a multitask network with a CNN as backbone and a parameters-free SVD-based head. This network can simultaneously predict subpixel coordinates of alignment marks’ keypoints and registration angle. The network takes image patches and their corresponding indices in the original image as input, rather than the entire original image. This allows the network to focus only on the unobstructed keypoint regions. To train the network without manual annotation of subpixel keypoints, a multicomponent loss function based on angular relationships is introduced. Graphical abstract