SoulNet: ultrafast optical source optimization utilizing generative neural networks for advanced lithography

Abstract

Abstract. An optimized source has the ability to improve the process window during lithography in semiconductor manufacturing. Source optimization is always a key technique to improve printing performance. Conventionally, source optimization relies on mathematical–physical model calibration, which is computationally expensive and extremely time-consuming. Machine learning could learn from existing data, construct a prediction model, and speed up the whole process. We propose the first source optimization process based on autoencoder neural networks. The goal of this autoencoder-based process is to increase the speed of the source optimization process with high-quality imaging results. We also make additional technical efforts to improve the performance of our work, including data augmentation and batch normalization. Experimental results demonstrate that our autoencoder-based source optimization achieves about 105  ×   speed up with 4.67% compromise on depth of focus (DOF), when compared to conventional model-based source optimization method.