Fast inverse lithography based on dual-channel model-driven deep learning.

Area of Science:

• Semiconductor Manufacturing
• Computational Lithography
• Deep Learning Applications

Background:

• Inverse lithography technology (ILT) is crucial for compensating image distortion in optical lithography.
• High computational complexity and data volume pose significant challenges for traditional ILT algorithms.
• Existing methods struggle to balance computational efficiency with high image fidelity.

Purpose of the Study:

• To introduce a novel dual-channel model-driven deep learning (DMDL) method for ILT.
• To address the computational burden and improve image fidelity in optical lithography.
• To develop an efficient and accurate ILT solution for semiconductor manufacturing.

Main Methods:

• A dual-channel model-driven deep learning (DMDL) network architecture derived from inverse optimization.
• Integration of a dual-channel structure for simultaneous mask contour modification and sub-resolution assist feature insertion.
• Implementation of an unsupervised training strategy using an auto-decoder to eliminate manual labeling.

Main Results:

• The DMDL method demonstrates superior computational efficiency compared to state-of-the-art ILT algorithms.
• DMDL achieves significantly improved image fidelity on semiconductor wafers.
• The dual-channel architecture effectively enhances the capacity for mask pattern optimization.

Conclusions:

• The proposed DMDL method offers a breakthrough in ILT by overcoming computational limitations and enhancing image fidelity.
• DMDL provides a more efficient and accurate solution for photomask pre-warping in semiconductor lithography.
• This deep learning approach represents a significant advancement for next-generation optical lithography systems.