Metrology and Defect Inspection

Lesson 5's control charts and capability indices both depend on having actual measured data to plot. This lesson covers where that data, specifically defect data, comes from: the inspection equipment that scans a wafer for anything unexpected, and the classification process that turns a raw detected anomaly into information useful for process control and, in lessons 7-8, yield learning.

Reading time: about 8 minutes.

6.1 Optical inspection: brightfield and darkfield

Most wafer inspection, especially high-volume, every-lot or every-few-lots inspection, is optical, because optical inspection is fast relative to the alternative, electron-beam inspection (6.2), even though it has a resolution floor set by the same diffraction limits the Lithography, Etch & CMP path's lesson 3 covered for lithography. Two complementary optical inspection modes are used:

• Brightfield inspection illuminates the wafer and collects the directly reflected light, forming an image closer to how the human eye would see the surface; a defect shows up as a region where reflected intensity or pattern differs from its surroundings or from a reference (either a neighboring, nominally identical die, die-to-die comparison, or a stored reference image, die-to-database comparison).
• Darkfield inspection instead collects only light scattered at an angle away from the direct reflection path, with the directly reflected beam deliberately excluded from the detector. A flat, defect-free surface scatters comparatively little light in these off-angle directions, so darkfield images are mostly dark except where something, a particle, a scratch, an edge or bump, scatters light toward the detector. Darkfield is often more sensitive to small particles and certain surface topography defects than brightfield, precisely because it is not competing against the much brighter directly reflected signal a flat surface returns.

Many inspection tools offer both modes, and a defect's appearance (or lack of appearance) in each mode is itself a useful clue toward what kind of defect it actually is, before any further review.

Inspection sensitivity (the smallest defect size reliably detected) and inspection speed (area scanned per unit time) trade directly against each other for a given optical wavelength and tool configuration, the same resolution-versus-practicality trade-off this section's lithography lessons already described for patterning, here applied to measurement instead. A fab therefore does not run every inspection at maximum sensitivity on every wafer; inspection recipes are tuned per layer, with the most defect-critical, least forgiving layers (where even a small defect is likely to be a killer defect) inspected more frequently and at higher sensitivity than more forgiving layers, balancing total metrology cost and cycle time impact against the actual yield risk each layer carries.

6.2 Electron-beam inspection and review

E-beam (electron-beam) inspection scans a focused electron beam across the wafer, similar in principle to CD-SEM (Lithography, Etch & CMP lesson 6.2) but used here for defect detection across a wider area rather than measuring a single critical dimension at a specific site. E-beam inspection resolves far smaller defects than optical inspection can, since its effective resolution is not limited by visible or near-visible light's wavelength the way optical inspection is, but it is correspondingly much slower, scanning an area at electron-beam speed rather than an optical system's much faster parallel imaging. For this reason, e-beam inspection is typically used selectively, on a sampled subset of wafers or specific high-risk layers, rather than on every wafer of every lot the way fast optical inspection commonly is.

Defect review is a related, separate step: once optical (or e-beam) inspection flags defect locations across a wafer, a review tool, often a higher-resolution SEM, revisits each flagged location individually to capture a detailed image and help classify what the defect actually is, since the original detection scan is optimized for speed and coverage, not for the detailed imaging review needs.

6.3 Defect classification

A detected and reviewed defect is classified, sorted into categories based on its appearance, likely cause, and likely electrical impact. Classification can be manual (an engineer visually reviewing the image) or automated (automatic defect classification, ADC, using image-recognition software, increasingly machine-learning-based, trained on previously classified examples to sort new defects by category without a person reviewing every single one). Typical categories distinguish, among others: particle contamination (a foreign object landed on the wafer), a pattern defect (an actual error in the patterned feature itself, a bridge between two lines that should not be connected, or a line that is open where it should be continuous), and a scratch or other mechanical damage. This classification feeds two different uses directly: process control (which tool, which step, is generating this kind of defect, feeding back into the SPC and FDC methods lesson 5 covered) and yield learning (lesson 8 covers using classified defect data, mapped across many wafers, to find the dominant yield-limiting mechanism).

A Pareto chart, ranking defect categories by how many total defects (or how much yield loss) each category accounts for, is a common first step once enough classified data has accumulated: it is rarely worth a yield team's attention to chase every low-frequency defect category with equal priority, and a Pareto view quickly shows which one or two categories, often called the "top offenders," are responsible for the bulk of current yield loss, directing limited engineering attention toward the categories where fixing the underlying cause will move the yield number the most.

6.4 Interview questions

What is the main difference between brightfield and darkfield optical inspection? Brightfield collects the directly reflected light and sees a defect as a contrast change; darkfield excludes the direct reflection and collects only off-angle scattered light, so a defect appears as a bright spot against an otherwise dark background, often giving better sensitivity to small particles.

Why is e-beam inspection used selectively rather than on every wafer? It resolves much smaller defects than optical inspection but is far slower, since it scans with a focused electron beam rather than optical inspection's faster, more parallel imaging; the throughput cost makes full-wafer, every-lot e-beam inspection impractical for most use cases.

What is defect review, and how does it differ from the initial inspection scan? A separate step where a higher-resolution tool, often a review SEM, revisits each location flagged by the faster initial inspection scan, capturing detailed images to support classification; the initial scan is optimized for speed and coverage, not detailed imaging.

What is automatic defect classification (ADC)? Image-recognition software, increasingly machine-learning-based, that sorts detected defects into categories automatically, trained on previously classified examples, reducing the need for an engineer to manually review every single defect image.

Name two downstream uses of classified defect data. Process control, identifying which tool or step is generating a particular defect type and feeding back into SPC/FDC, and yield learning, using defect data mapped across many wafers to find the dominant yield-limiting mechanism.