Inverse optical scatterometry using sketch-guided deep learning.

Area of Science:

• Nanotechnology
• Metrology
• Artificial Intelligence

Background:

• Optical scatterometry (OS), or optical critical dimension (OCD) metrology, is vital for semiconductor nanostructure characterization.
• Model-based OS involves solving complex inverse problems, often limited by predefined geometric models.
• Existing methods, including deep learning, struggle with generalizability due to model constraints.

Purpose of the Study:

• To introduce a novel sketch-guided neural network (SGNN) for nanostructure reconstruction in optical scatterometry.
• To overcome the limitations of predefined geometric models in inverse scattering problems.
• To develop a more accurate and generalizable method for semiconductor metrology.

Main Methods:

• Developed a sketch-guided neural network (SGNN) trained on a generic profile model.
• The SGNN learns scattering principles and sketching techniques to reconstruct nanostructure profiles from optical signatures.
• Validated the approach using one-dimensional gratings and compared performance against existing methods.

Main Results:

• The SGNN accurately reconstructs nanostructure profiles, demonstrating strong generalizability beyond the training model.
• Performance is comparable to nonlinear regression and superior to traditional deep learning methods.
• This marks the first application of sketching concepts in deep learning for inverse scattering problems.

Conclusions:

• The SGNN offers a novel solution for nanostructure reconstruction in optical scatterometry.
• The method enables fast and accurate characterization of nanostructures in semiconductor manufacturing.
• This approach enhances the applicability and robustness of optical metrology techniques.