Fast inverse lithography approach based on a model-driven graph convolutional network.

Area of Science:

• Semiconductor Manufacturing
• Computational Lithography
• Machine Learning in Engineering

Background:

• Advanced optical lithography systems rely on inverse lithography techniques (ILT) to optimize photomask transmission functions for improved image fidelity.
• Traditional ILTs face computational intensity challenges, hindering their widespread adoption in high-volume integrated circuit manufacturing.

Purpose of the Study:

• To develop a computationally efficient and high-fidelity ILT framework for advanced optical lithography.
• To address the limitations of traditional ILTs in terms of speed and resource requirements for integrated circuit manufacturing.

Main Methods:

• A model-driven graph convolutional network (MGCN) framework is proposed, integrating dense concentric circular sampling (DCCS) for feature extraction.
• A GCN-based encoder predicts optimized mask patterns, followed by a model-driven decoder that simulates the lithography imaging process.
• An unsupervised training strategy is employed to eliminate the need for time-consuming labeled data.

Main Results:

• The MGCN framework, leveraging DCCS and an unsupervised approach, significantly improves computational efficiency for ILT.
• The proposed method achieves high-fidelity imaging results, comparable to or exceeding state-of-the-art ILT methods.
• Fast mask prediction is enabled through parallel computing on a GPU framework.

Conclusions:

• The MGCN approach offers a viable solution for accelerating ILT in high-volume manufacturing of integrated circuits.
• This framework demonstrates superior performance in both computational speed and imaging fidelity compared to existing ILT methods.
• The integration of machine learning with lithography process modeling presents a promising direction for future advancements.