Detection limits of AI-based SEM dimensional metrology

The speed of in-line scanning electron microscope (SEM) measurements of linewidth, contact hole, and overlay is critically important for identifying the measurement area and generating indispensable process control information. Sample charging and damage to sensitive integrated circuit structures require low primary electron beam current and dose. These demands inevitably result in noisy, low-contrast images, which (depending on the noise and contrast) can make traditional dimensional measurements no longer viable. Object segmentation based on artificial intelligence (AI) models can extend the viability of these measurements, but misleading, wrong results may occur due to false positives and false negatives with respect to the ground truth information. This work establishes the methodology for quantifying detection limits for AI-based measurements from SEM images. The detection limits establish the relationship between the quality of SEM images and human or AI model detection performance.