SWIR wafer inspection: seeing subsurface cracks silicon hides

A polished wafer can pass a visible-light inspection with a mirror finish and still hide cracks, voids and residual layers beneath the surface. Those flaws grow through later high-temperature steps and surface at dicing, killing dies that already carry weeks of accumulated value. This guide is for process and yield engineers choosing a camera for incoming-wafer, post-grind and dicing-lane checks. It explains why silicon turns transparent just past 1100 nm, how to pick the wavelength, how a line scan SWIR camera covers a 300 mm wafer, and how to size optics and backlight, with the arithmetic shown and a parts list at the end.

Why can a SWIR camera see cracks that visible light misses?

Silicon stops absorbing where its bandgap ends: 1.12 eV corresponds to 1107 nm, so above roughly 1150 nm a bare wafer transmits part of the light and an indium gallium arsenide sensor records the crack that scatters it.

A visible camera cannot follow, because a silicon sensor goes blind at silicon's own absorption edge — the detector and the wafer are the same material. InGaAs sensors respond from 900 nm out to 1700 nm, which brackets the whole working window.

In transmission the geometry is simple: light enters the backside, crosses the wafer, and a subsurface void displaces silicon and scatters the beam, so the crack reads as a dark line on a bright background. The SWIR, NIR or visible band comparison shows where this window sits against its neighbours.

Which wavelength should you tune the inspection to?

Work between 1200 and 1350 nm for a full-thickness polished wafer, and push to 1500–1600 nm when the stack is thicker or more absorbing; design around the 1380–1450 nm dip where water vapour in the beam path swallows signal.

Absorption falls steeply above the 1107 nm edge, so every extra 100 nm of wavelength buys transmission; the price is resolution, which scales linearly with wavelength. Staying at 1300 nm rather than 1550 nm keeps about 20% more spatial detail, because 1550 / 1300 ≈ 1.19. The water band matters even over the short air path on a bench, so treat 1350–1450 nm as a dead zone and jump across it.

The material-by-material logic in SWIR wavelength selection by material extends to silicon, and near infrared wavelength selection for water and silicon inspection covers the water and silicon cross-over in detail.

Do you need line scan or area scan for a 300 mm wafer?

Match the line length to the field you actually need to resolve: 1024 pixels across the full 300 mm gives 293 μm per pixel — fine for breakage and void screening — while 25 μm per pixel across 300 mm would need 12,000 pixels, so fine-crack work becomes band scanning.

Ground sample distance is field width divided by pixels, and the smallest crack you can catch is about 2 pixels wide:

The decision is a scan plan: run the full wafer at 293 μm per pixel for incoming screening, then re-scan the edge exclusion ring and the dicing streets inside a 25–40 mm band. If the wafer sits still instead of moving, the area scan camera category filters by SWIR spectrum, and the line scan camera category holds the moving-wafer options. The wafer line scan resolution guide runs the same arithmetic backwards from defect size.

For the band-scan geometry, this 1K GigE SWIR line is the baseline:

How much light gets through, and which illuminator survives?

Expect roughly half the light to survive: bare silicon reflects about 30% at each surface, which caps a plain wafer near 50% throughput before absorption, so the backlight must be a broadband SWIR source rather than a visible LED.

The 30% comes straight from Fresnel reflection at silicon's refractive index of about 3.5, and two uncoated faces in series leave roughly half: 0.7 × 0.7 ≈ 0.5. An AR-coated reference wafer buys much of that back, so keep one in the drawer for set-up and contrast checks.

Tungsten-halogen lamps deliver broadband output well past 1600 nm. Near-infrared LEDs come at discrete wavelengths, typically 1200, 1310, 1450 and 1550 nm, so pick one outside the water dip. Ordinary machine-vision LED lighting fades out near 1050 nm and is useless here. Put a diffuser under the wafer for an even background; the fixes in the line scan illumination uniformity guide apply directly, and backlight wavelength selection for semi-transparent parts covers the transmission geometry.

What focal length and working distance does the set-up need?

Size the lens from the pixel, not the wafer: 25 μm ground sample distance on 12.5 μm pixels is 0.5× magnification, a 150 mm working distance then needs a 50 mm lens, and the 1024-pixel line covers 25.6 mm of silicon.

Worked example, both rows of the table above: at 25 μm GSD, m = 12.5 / 25 = 0.5, FOV = 1024 × 25 μm = 25.6 mm, and f = 150 mm × 0.5 / 1.5 = 50 mm. For the 40 mm band at 39 μm, m = 12.5 / 39 ≈ 0.32 and f = 150 mm × 0.32 / 1.32 ≈ 36 mm, so a 35 mm lens is the nearest standard. The lens focal length calculator runs the same numbers from any other working distance.

Order SWIR-corrected optics: plain glass focuses 1300 nm at a different distance from the visible alignment image, so a lens that looks sharp on the alignment pass can still be soft at the working wavelength. The diffraction floor also rises with wavelength — Airy diameter = 2.44 × λ × f-number, which at 1300 nm and f/4 gives 2.44 × 1.3 μm × 4 ≈ 12.7 μm, about one pixel of blur. The SWIR lens basics guide explains the glass problem, and line scan lens selection for 4K, 8K and 16K cameras covers resolution across a line.

How fast can you scan before the signal falls apart?

Line rate is never the bottleneck: at 25 μm per pixel a 50 mm/s chuck feed asks for 2,000 lines per second, slow for any line scan camera, so the real limits are exposure per line and how many photons the backlight pushes through the silicon.

The arithmetic: line rate = feed speed ÷ GSD = 50 mm/s ÷ 0.025 mm = 2,000 Hz. Exposure per line cannot exceed the line period, so at 2,000 Hz you get up to 500 μs of integration; if the histogram runs thin, slow the stage or add light rather than stretching exposure. The LS-LCM-0038 tops out at 29.4 kHz. That ceiling would carry the same 25 μm GSD to 735 mm/s — a speed the stage motor reaches long before the photons run out.

Trigger one line per GSD of travel from the stage encoder so pixels stay square; encoder ticks per pixel for line scan sets the divider, and encoder-triggered line scan capture: wiring quadrature inputs covers the wiring. For the exposure and motion blur trade, camera fps and shutter speed for moving web inspection is the reference.

Which SWIR line scan camera fits which interface?

Both cameras share the same 1K InGaAs line, 12.5 μm pixels and M42 mount; pick the GigE model for a standard network port up to 29.4 kHz, or the Camera Link model for an existing frame grabber and the 40 kHz ceiling.

All 2 share: Pixel size 12.5 μm; Resolution 1024 x 1; Lens mount M42. Values as printed in each product's data sheet; open a part for its drawing and full specifications.

The Camera Link version also offers a 950–1700 nm band option, which trims out-of-band flare when the illuminator leaks short wavelengths.

For a rack that already has a Camera Link frame grabber:

An uncooled InGaAs line drifts with ambient temperature, so capture a flat-field reference at the start of each shift and after warm-up; uncooled SWIR camera stability for inspection systems quantifies the drift and the countermeasures. Filter the camera finder by SWIR spectrum and line length to see every alternative, or start from your GSD in the camera configurator. When the project grows from a bench camera into a whole inspection cell, the system-level trade-offs are laid out in this industrial vision system selection guide for silicon wafer inspection.

A first bench set-up then needs five parts:

• 1K SWIR line scan camera, GigE or Camera Link to match the grabber
• SWIR-corrected 50 mm lens in M42 for the 25 μm band, or 35 mm for the 40 mm band
• Tungsten-halogen or 1300–1550 nm LED backlight with diffuser and shroud
• Stage encoder divided down to one trigger per GSD
• AR-coated reference wafer for flat-field and contrast checks

Where does SWIR transmission inspection stop working?

Transmission ends where bare silicon does: metal layers and filled vias are opaque, heavily doped regions absorb through free-carrier effects, a ground-damaged backside scatters the beam, and bonded multi-die stacks block it outright — so keep a visible channel for surface defects.

Heavily doped substrates raise free-carrier absorption across the whole band, so a wafer that transmits brightly when lightly doped can go nearly opaque after a heavily doped backside treatment; check the resistivity before promising numbers. Backgrind damage roughens the exit face and scatters light before it forms an image, which polish or coating restores. InGaAs is monochrome, so colour and surface shape stay a visible job — pair a silicon line scan camera with the SWIR line, or read why UV imaging reveals surface defects visible light cannot for the other end of the spectrum.

The decision, case by case: incoming bare-wafer screening runs the full 300 mm at 293 μm per pixel under a halogen backlight at 1200–1300 nm. Dicing-street and edge monitoring re-scans a 25–40 mm band through a 35–50 mm lens at 25–39 μm per pixel, with a 50–78 μm crack floor. Thinned wafers start nearer 1200 nm. Patterned, bonded or heavily doped stacks move to visible reflection or X-ray instead of transmission. More selection guides sit on the cameras topic page, and the applications team will size the line, GSD and lens from your wafer spec under the free expert support that comes with every order — start at the contact page if the configurator does not settle it.