Curvilinear Sub-Resolution Assist Feature Placement Through a Data-Driven U-Net Model.

Area of Science:

• Semiconductor Manufacturing
• Computational Lithography
• Deep Learning Applications

Background:

• Sub-resolution assist features (SRAFs) are critical in advanced semiconductor manufacturing for process window enhancement.
• Conventional SRAF placement methods face speed-pattern fidelity trade-offs and struggle with complex layouts for advanced nodes.

Purpose of the Study:

• To develop a deep learning framework for accelerating and improving the optical refinement of SRAF shapes.
• To overcome the limitations of traditional SRAF placement methodologies in terms of speed and fidelity for complex layouts.

Main Methods:

• A large-scale dataset was created using coarse, binarized SRAF patterns as input.
• Ground-truth SRAF shapes were generated using a Level-Set Method (LSM) optimized for optical performance.
• A U-Net convolutional neural network was trained to map coarse SRAF inputs to optically optimized outputs.

Main Results:

• The deep learning model achieved multi-order-of-magnitude acceleration compared to traditional CPU-based methods.
• The U-Net model demonstrated significantly faster performance than GPU-accelerated algorithms while maintaining high pattern fidelity.
• Optically optimized SRAFs generated by the U-Net showed high fidelity to ground-truth layouts and comparable optical performance.

Conclusions:

• A data-driven surrogate model can effectively replace traditional algorithms for SRAF optical refinement.
• This deep learning approach offers a promising solution to mitigate computational costs in mask synthesis.
• The findings provide a foundation for future integrated optimization solutions in computational lithography.