Sufficient machine learning based pattern classification for curvilinear designs

In this paper, we address a critical challenge in electronic design automation: the classification of curvilinear patterns with mirror and rotational symmetry. Traditional pattern classification tools are inadequate for curvilinear layouts due to their sensitivity to arbitrary edge angles, creating significant limitations for semiconductor manufacturing, silicon photonics, AR/VR, MEMS and biomedical device industries.

To solve this problem, we developed a comprehensive six-step machine learning approach. This methodology begins by designing feature vectors that can differentiate directions, followed by defining Points of Interest from verification checks. The process continues with capturing and analyzing feature vectors at these points, training machine learning models and then applying these models for pattern classification.

We validated our approach through four test cases using ellipse arrays in both isolated and dense configurations at various orientations, as well as with a real-world curvilinear Mask Process Correction verification layout. Results demonstrated that the methodology successfully classified patterns with mirror and rotational symmetry, correctly identified similar patterns regardless of orientation and properly distinguished patterns based on environmental context rather than orientation alone. Most importantly, it proved effective when applied to complex curvilinear layouts in semiconductor manufacturing.

This method enables more efficient pattern classification in curvilinear layouts across multiple high-tech industries. By reducing engineering review time, facilitating pattern databank construction and improving overall work efficiency, this methodology provides a valuable solution for industries increasingly reliant on curvilinear designs for advanced applications.

This paper was originally presented at the 2024 SPIE Advanced Lithography + Patterning conference.

Lianghong Yin, Marko Chew, Shumay Shang, Le Hong, Fan Jiang, Ingo Bork, Ilhami Torunoglu, "Sufficient machine learning-based pattern classification for curvilinear layouts," Proc. SPIE 13216, Photomask Technology 2024, 132162R (20 November 2024); https://doi.org/10.1117/12.3034740

What you'll learn:

• How machine learning can overcome traditional pattern classification limitations by effectively handling mirror and rotational symmetry in curvilinear layouts
• A six-step methodology for implementing ML-based pattern classification that focuses on environmental context rather than orientation
• Practical applications of this technology across multiple industries including semiconductor manufacturing, Silicon Photonics, AR/VR, MEMS and biomedical devices

Who should read this:

• Electronic Design Automation (EDA) engineers and researchers developing verification tools for advanced semiconductor manufacturing
• Machine learning specialists working on computer vision and pattern recognition applications in manufacturing
• Semiconductor industry professionals involved in Optical Process Correction (OPC) and Mask Process Correction (MPC)
• R&D engineers in Silicon Photonics, AR/VR, MEMS and biomedical device industries dealing with curvilinear layouts
• Technical managers and decision-makers evaluating advanced verification methodologies for next-generation manufacturing processes