Semiconductor wafer inspection: line scan resolution from defect size

A mirror-finish 300 mm wafer hides its problems well: particles a fraction of a micron across, scratches buried in specular glare, pattern defects inside dense die rows. If you are sizing a line scan camera, lens and light for semiconductor wafer inspection — alignment, defect zones or die-level review — this guide gives you the arithmetic. You will compute the magnification a defect size demands, the field of view one scan line covers, how many tracks cross the wafer, the line rate at your stage speed and the light that magnification costs, then decide which zones a line scan station can inspect and where it must hand the job over.

How many pixels does a wafer defect need?

Three pixels across a defect's smallest dimension is the working minimum for reliable detection, and five or more once you must size or classify it, so a 0.5 µm particle needs an object-side pixel of at most 0.17 µm. Three pixels is where a defect produces a shape your software can threshold and hold against noise; five pixels give it a profile you can measure and classify. The same pixel rule that sizes defect cameras in general applies on silicon, and when the defect must be measured rather than merely found, the micron-level semiconductor resolution guide works through the tighter budget.

What magnification does your smallest defect demand?

Magnification obeys m = 3 × camera pixel ÷ smallest defect: 3 µm defects need 3.5× with 3.5 µm pixels, 0.5 µm particles need 21×, and 30 µm defects need 0.7× with 7 µm pixels.

Worked example, die-level zone on a 16K, 3.5 µm sensor, target 3 µm defects:

• Magnification: m = 3 × 3.5 / 3 = 3.5×. The nearest catalog lens family works at 3.35×, so the design carries 3.35×.
• Object pixel: 3.5 / 3.35 = 1.04 µm, so the smallest reliably detected feature is 3 × 1.04 = 3.1 µm — the target met with nothing to spare.
• Sensor length: 16,384 × 3.5 µm = 57.3 mm, and line FOV = 57.3 / 3.35 = 17.1 mm.
• The same 3.35× on 5 µm pixels: object pixel 5 / 3.35 = 1.49 µm, detection floor 3 × 1.49 = 4.5 µm, sensor 16,384 × 5 µm = 81.9 mm, line FOV 81.9 / 3.35 = 24.5 mm.

Both sensor lengths fit the image circle this lens class provides — the check that matters before ordering, because an undersized circle dims the ends of a long line, as the vignetting guide explains. For the trade space across 4K, 8K and 16K optics, the line scan lens selection guide is the companion read. The lens that carries the 3.35× case:

How much of a 300 mm wafer does one scan line cover?

Line field of view is sensor length divided by magnification: a 16K, 3.5 µm sensor at 3.35× covers 17.1 mm, so 18 indexed tracks cross a 300 mm wafer, while 21× magnification shrinks the line to 2.7 mm and needs 110 tracks.

Every row is the same two divisions — sensor length ÷ magnification for the strip, speed ÷ object pixel for the rate — and the line scan field-of-view guide derives them step by step. For the alignment row the line rate is a non-issue: the task asks 1.4 kHz. The LS-LCM-0016 — 2K, 7 µm, GigE, TDI, M42 — runs to 58 kHz, so the choice there is optics, not electronics.

What line rate keeps the scan sharp at wafer speed?

Line rate equals stage speed divided by object pixel: at 50 mm/s with a 1.04 µm object pixel the camera must fire about 48 kHz, and any mismatch between them smears the image by exactly the missing fraction.

The arithmetic: 50 mm/s is 50,000 µm/s, and 50,000 / 1.04 = 48,077 lines per second, call it 48 kHz. Each line then owns 1 / 48,000 s = 20.8 µs of exposure, and that window is the whole light budget. The line rate walkthrough covers the camera side, encoder ticks per pixel ties every line to real travel, and periodic speed ripple with no encoder fitted has its own jitter correction guide.

Where line rate outruns the light, time-delay integration is the standard answer: charge shifts along the sensor in step with the motion, stacking stages of light collection without changing the geometry:

A four-stage TDI sensor collects about four times the charge per pixel that a single line collects — the margin that rescues a 48 kHz scan from a light-starved strip.

How much light does high magnification cost?

Effective f-number grows as F# × (1 + m), so an F4.5 lens working at 3.35× behaves like F19.6 and delivers about 19 times less light — near 4 stops — while 21× magnification costs 484 times, close to 9 stops.

The loss factor is (1 + m)²: 4.35² = 18.9 at 3.35×, and 22² = 484 at 21×. Brightness per unit area at the wafer must rise by the same factor, which is why high-magnification strips live or die on the light budget. What that means in watts and duty cycle on a fast line is the line scan light power question, and firing the light only inside the 20.8 µs window is what a line scan lighting controller is for.

Size the emitting length to cover the line FOV plus about 10%, so 18.8 mm here. For a die-level strip the usual pick is a high-output line light:

Its 300 × 17 mm emitting window and 87 W red / 108 W white output make it the reference part for this strip's light budget.

Which lighting geometry shows defects on a mirror-finish wafer?

The law of reflection decides: bright field reads the light a polished surface sends back into the lens, dark field reads only the scattered light, and scattered light is exactly what particles and scratches produce, so they show bright on black.

A wafer is a mirror. In bright field the surface returns the beam and reads white, so a sub-pixel particle must dim a bright background — it barely manages. In dark field the light arrives at an angle and only scatter reaches the lens, so the same particle glows against black and can be caught at fewer pixels per defect, which buys scan speed for a given defect size. The geometry trade is worked through in bright field versus dark field and the wider set of line scan illumination techniques.

Wavelength helps too: scattering from particles far smaller than the wavelength rises as 1/λ⁴, so a 395 nm light scatters roughly (660 / 395)⁴ = 7.8, about 8 times, more strongly than a 660 nm red one. Blue and near-UV line lights earn their keep on particle hunts, and light color versus monochrome contrast sets the pairing with the sensor. Whatever the geometry, uniformity along the strip decides whether clean areas read as false defects — the dark-band fix covers that failure.

What does a practical wafer inspection station include?

One station solves one zone well: an encoder-driven stage or gantry, a line scan camera sized to that zone's defect budget, a lens whose image circle covers the full sensor length, a strobed line light, and a controller that pulses per line.

Full coverage stacks zones — alignment optics on the edge, an overview pass near 0.75×, die-level strips at 3.35× where the process calls for them. The camera side spans line scan cameras from 1K to 16K, the optics line scan lenses by image circle and magnification, and the illumination line scan lights by length, width and wavelength; the light finder and camera finder filter the same fields.

Before the first line fires, the handling system has to prove a thin wafer is seated and flat: thin wafer presence sensors for semiconductor robot end-effectors are the parts built for that check. Wire the encoder into the camera or frame grabber, tie the strobe to the line period, and run one strip end to end before scaling to all tracks.

When is a line scan camera the wrong tool for a wafer?

Below roughly 1 µm object pixels the build stops closing: 21× magnification leaves a 2.7 mm line, 110 tracks, 9 stops of lost light and a 300 kHz line rate, so sub-micron whole-wafer contamination inspection belongs on dedicated dark-field inspection tools.

The data rate alone says it: 16,384 pixels × 300,000 lines/s ≈ 4.9 GB/s of raw image data before any processing. A line scan station earns its keep from 100 µm alignment details down to about 3 µm die-level defects; between that and sub-micron contamination sits purpose-built equipment with deep-UV optics and TDI arrays. For slow or static checks — a sample wafer on a lab stage — an area scan camera driven in line scan mode reuses the same arithmetic with less wiring, and when defect classes multiply, deep learning on wide 8K strips takes over classification after the optics resolve the defect.

Which checks settle the build before you order?

Five numbers close the decision: magnification against the 3-pixel floor, line FOV against track count, line rate against the camera's maximum, effective f-number against the light budget, and image circle against sensor length.

• Compute m = 3 × pixel ÷ defect per zone and pick the lens first; the lens sets what is possible, the camera then fits it.
• Multiply line FOV by tracks and confirm the sum crosses 300 mm with edge overlap; recount tracks at the real magnification, not the nominal one.
• Compute line rate = speed ÷ object pixel and keep 20-30% headroom to the camera's maximum line rate.
• Apply (1 + m)² to the light budget and check the strobe fits the line period — 20.8 µs at 48 kHz.
• Confirm the image circle covers the sensor length, then image one strip carrying known defects before ordering the rest of the motion.

Size the zones, then pick the parts: the camera configurator walks camera, lens and light in order, and if a zone lands past the 3 µm line, the contact page is where to draw the boundary with us.