GLAU-Net: a fast lithography imaging model for chemically amplified resists

Area of Science:

• Computational lithography
• Deep learning in semiconductor manufacturing

Background:

• Rigorous lithography simulation is crucial but computationally intensive.
• Exposure and post-exposure bake (PEB) are key steps influencing photoresist profiles.
• Chemically amplified resists are widely used in semiconductor fabrication.

Purpose of the Study:

• To develop a novel deep-learning model (GLAU-Net) for accelerating lithography simulation.
• To improve the efficiency of exposure and PEB steps in chemically amplified resist simulation.
• To enhance the accuracy of latent image and photoresist profile prediction.

Main Methods:

• Proposed a deep-learning model named GLAU-Net.
• Incorporated a mask attention gate module for critical mask regions.
• Utilized a composite loss function (MSE + gradient loss) for boundary accuracy.

Main Results:

• GLAU-Net achieved 235x speedup over CPU-based rigorous models.
• GLAU-Net achieved 4x speedup over GPU-accelerated rigorous models.
• Outperformed existing deep learning models in latent image and photoresist profile prediction accuracy.

Conclusions:

• GLAU-Net offers a significant acceleration for lithography simulations.
• The model demonstrates superior accuracy in predicting latent images and photoresist profiles.
• GLAU-Net is a promising tool for optimizing semiconductor manufacturing processes.