Label-free neural networks-based inverse lithography technology

Area of Science:

• Semiconductor Manufacturing
• Computational Lithography
• Artificial Intelligence in Engineering

Background:

• Traditional inverse lithography technology (ILT) mask optimization is computationally intensive.
• Neural network-based ILT (NNILT) offers improved efficiency but relies heavily on labeled data.
• Label dependency limits NNILT's performance and extrapolation capabilities for mask optimization.

Purpose of the Study:

• To develop a label-free NNILT (LF-NNILT) for advanced photolithography mask optimization.
• To enhance printability of target layouts and manufacturability of synthesized masks.
• To significantly accelerate the mask optimization process.

Main Methods:

• Implementation of a novel label-free neural network architecture for inverse lithography.
• Training and validation of LF-NNILT without relying on pre-existing labeled datasets.
• Comparative analysis against traditional ILT and label-dependent NNILT methods.

Main Results:

• LF-NNILT demonstrates superior printability and mask manufacturability compared to traditional ILT.
• The optimization speed of LF-NNILT is two orders of magnitude faster than traditional ILT.
• LF-NNILT achieves high performance and extrapolation ability without label dependency.

Conclusions:

• Label-free NNILT is a viable and highly efficient approach for advanced photolithography mask optimization.
• LF-NNILT simplifies implementation and offers improved solvers, supporting next-generation lithography development.
• This method overcomes the limitations of label-dependent NNILT, paving the way for faster and more effective mask design.