On the feature accuracy of deep learning mask topography effect models

A deep-learning-based lithography model using a generative neural network (GAN) approach is developed and assessed for its ability to predict aerial images at different resist heights. The performance of the GAN approach is evaluated by analyzing deviations between model-generated aerial images and golden images, as well as differences in critical dimension (CD) values. Additionally, error analysis is conducted based on the feature distribution of each photomask. Selected patterns and their aerial images are compared both qualitatively to assess local errors and quantitatively through root-mean-square (RMS) errors to evaluate global accuracy. Error analysis reveals the features produced by the deep learning model leading to the highest deviation from the rigorous model results, and the error is decomposed into the error contributions of underpredicted and over- predicted features. An array of aerial images for selected resist heights produced by the deep learning model is assessed, revealing increasing errors with increasing resist heights. The limitations of applying deep learning techniques in computational lithography are illustrated by comparing a target pattern with and without optical proximity correction (OPC) features.