Model-driven convolution neural network for inverse lithography.

Area of Science:

• Semiconductor Manufacturing
• Computational Imaging
• Machine Learning

Background:

• Optical lithography is crucial for integrated circuit fabrication but suffers from image distortions.
• Inverse lithography techniques (ILT) mitigate these distortions, enhancing resolution and fidelity.
• Increasing integrated circuit density presents computational challenges for conventional ILT methods.

Purpose of the Study:

• To develop a novel framework using a model-driven convolution neural network (MCNN) to address the computational complexity of ILT.
• To provide an approximate solution for ILT that significantly reduces the number of iterations required by traditional algorithms.
• To enhance the speed and imaging performance of optical lithography systems.

Main Methods:

• A model-driven convolution neural network (MCNN) framework was developed, leveraging the imaging model of optical lithography.
• The neural network architecture and parameters were derived from an unfolded and truncated model-based iterative ILT procedure.
• An unsupervised training strategy was employed using a model-based decoder, eliminating the need for time-consuming data labeling.

Main Results:

• The proposed MCNN framework, when combined with gradient-based methods, accelerates ILT optimization by up to an order of magnitude.
• The approach demonstrates considerable performance advantages and improves imaging performance in coherent optical lithography.
• Simulation results verify the superiority and effectiveness of the MCNN approach for ILT problems.

Conclusions:

• This study introduces the first application of MCNN to solve the ILT problem, offering significant computational and performance benefits.
• The MCNN framework provides an efficient and effective method to improve the computational efficiency of ILT algorithms.
• This work opens new possibilities for MCNN techniques in advancing semiconductor fabrication technologies.