Relating human and AI-based detection limits in scanning electron microscopy dimensional metrology.

Area of Science:

• Metrology
• Artificial Intelligence
• Semiconductor Manufacturing

Background:

• Scanning Electron Microscopy (SEM) is crucial for nanoscale measurements in semiconductor manufacturing.
• Low electron beam current and dose in SEM imaging produce noisy, low-contrast images, hindering traditional analysis.
• Sensitive integrated circuit (IC) structures require careful handling to prevent charging and damage.

Purpose of the Study:

• To investigate the detection limits of AI-based SEM image segmentation for IC structure detection.
• To establish the relationship between SEM image quality and AI model accuracy.
• To compare AI model detection limits with human detection limits.

Main Methods:

• Utilized SEM image simulation software to generate datasets with varying noise and contrast.
• Characterized SEM images using 25 image quality metrics.
• Trained and evaluated three AI models on these image datasets.
• Mapped image quality characteristics to AI model accuracy metrics.

Main Results:

• Established detection limits for AI models based on image quality and confidence levels.
• Related AI model detection limits to human detection limits, using Rose's signal-to-noise ratio (SNR=5) as a benchmark.
• Demonstrated upper and lower SNR bounds for three AI models relative to human detection limits.

Conclusions:

• Developed a method to determine AI-based SEM dimensional metrology detection limits.
• The findings are relevant for semiconductor vendors and AI model consumers, enhancing trust in critical dimension measurements.
• The method allows for trusted AI model characterization for semiconductor production using noisy SEM images.