Proximity effect correction in lithography using a U-net neural network

Inverse design method in nanophotonic devices has been expanding its applications in various fields rapidly with the help of rapid development of artificial intelligence (AI). However, AI-based inverse designs often require more complex and high-density patterns compared to traditional rule-based designs. Therefore, the proximity effect that refers to the discrepancy between the designed patterns and the actual fabricated patterns in lithography processes gets worse because of the randomness in design. In random structures, a proximity effect correction (PEC) that makes this discrepancy correct requires complex dose calculations for each shape. We implemented PEC based on a model using a U-net structured neural network to reduce fabrication errors of AI-based designs with random and high-density patterns. By comparing PEC-applied patterns with non-PEC-applied patterns against the target design and evaluating quantitatively, we experimentally confirmed the usefulness of PEC based on a U-net neural network for AI-based designs.