Skip to content

THE COURSE / M9 MEASURE AND YIELD / M9.5

Finding defects

After this lesson you can:

  • Explain how an optical inspector finds defects by comparing dies.
  • Say why it can detect a defect far smaller than its light but not show what it is.
  • Weigh sensitivity against nuisance in setting an inspection, and read a defect map.

Loading the lesson…

Sources

  1. W. Kong, W. Zhang, L. Qin and D. Gong, "PWDD-Net: A Patterned Wafer Defect Detection Network for Semiconductor Manufacturing," Nanomaterials 16, 1026 (2026). Patterned-wafer inspection compares each die with a defect-free reference (die-to-reference) or with adjacent dies (die-to-die); the comparison is sensitive to sub-pixel registration errors and illumination changes.
  2. R. Zhang, X. Guo, Y. Xu, T. Zhang and M. Ming, "BWD-DETR: A Robust Framework for Bright-Field Wafer Defect Detection," Sensors 26, 1064 (2026). Optical microscopes image large areas fast; bright-field resolution is limited by diffraction, so small defects can't be imaged clearly and their signals are easily masked by noise; detection compares die to die, cell to cell or die to database.
  3. KLA, "Chip Manufacturing 101" (2021, read 25 October 2026). Defect sizes are often smaller than 10 nanometres; more than a thousand process steps for every chip; process control as inspection, metrology and data analysis.
  4. U. Amato, A. Antoniadis, I. De Feis et al., "Detecting Important Features and Predicting Yield from Defects Detected by SEM in Semiconductor Production," Sensors 25, 4218 (2025). STMicroelectronics production data: on average 2.3–2.4 of 30–35 layers inspected per wafer; in one product 4.7% of dice failed electrical test overall, 6.8% of those with a generic defect and 16.3% of those with an SEM-classified defect (9.1%, 13.2% and 19.0% in another); SEM with AI software estimates each defect's type.
  5. MIT OpenCourseWare, 2.830J/6.780J/ESD.63J Control of Manufacturing Processes, lecture 10, "Yield Modeling" (Spring 2008), after May and Spanos, CC BY-NC-SA. Poisson: Y = e^(−A₀D₀) for random, uncorrelated point defects that each cause a fault; negative binomial: Y = (1 + A₀D₀/α)^(−α), α the clustering parameter, Poisson as α grows large, typically 1.5 to 2 for memory and microprocessors; clustered defects land several to a die; critical area, and the defects that matter most near the minimum feature size.