You polish a part to mirror finish, and the inspection line rejects it for scratches no one can photograph. Dark field finds nothing because the scratch is shallower than the wavelength of the light and scatters almost nothing; coaxial light returns a perfectly uniform field. The fix is not more light, it is light that measures phase or slope. Differential interference contrast (DIC) and polarized pupil division convert nanometer-scale steps and fractions-of-a-degree tilt into grey levels an ordinary camera records. This explainer gives you the optics, the formulas with worked numbers, the field-of-view arithmetic, the camera settings, and a clear line between what standard lights and lenses can do and where you need an interference unit. It is written for engineers specifying mirror, wafer, sapphire and polished-metal inspection.
Why can't normal lighting see a nanometer scratch on a mirror?
A scratch 40 nm deep on a mirror finish scatters almost nothing, so dark field and coaxial bright field both return a blank: intensity lighting only sees features that scatter, and a smooth step or a gentle warp scatters nothing at any exposure.
A surface is optically smooth once its roughness falls below about λ/20, which is 26 nm at a green 525 nm wavelength. Mirror-polished metals, wafers and sapphire sit far below that mark, so nearly all the light leaves in the specular direction. A dark-field set-up catches only what scatters sideways, meaning ragged scratches from a few hundred nanometres upward, while a clean-walled step, polish texture and warp scatter essentially nothing at all.
Warp is the extreme case. A 0.1 degree bow across a 10 mm part is a 17 µm height change, more than 400 times the 40 nm step, and bright field shows none of it. If your defect list contains polish mark, orange peel, warp or nanometer step, geometry alone will not close the job. The law of reflection behind bright versus dark field and the placement of camera, light and part set how much light you collect, but both work in scattered or blocked light. Even beating specular reflection on copper only tames glare; none of it creates contrast from phase.
How does differential interference contrast turn height into brightness?
DIC splits polarized light into two beams that reflect from points a few micrometres apart on the part, and the optical path difference, doubled by reflection to twice the height difference, interferes into grey-level contrast at an analyzer.
The light path runs polarizer, a Nomarski-type prism that shears the beam into two polarizations separated by less than the diffraction spot (a fraction of a micrometre to a few micrometres at machine-vision magnifications), the lens, the part, back through the prism to recombine, and finally an analyzer. Both beams reflect from almost the same point; only the height difference between the two sampling points matters:
Because the two points sit a shear distance apart, Δh is the height change across that distance: the image is the height derivative along the shear axis. A step becomes a bright-dark line pair, a slope becomes even shading, and a feature running parallel to the shear axis produces no path difference and disappears. Rotate the part or the unit 90 degrees for a second pass; that single rule recovers most missed scratches.
How does polarized pupil division turn tilt into brightness?
Polarized pupil division sends polarized light across the whole lens pupil and reads how local surface tilt rotates it on reflection: a mirror tilted 0.5 degrees rotates linear polarization by a full degree, which an analyzer set just off extinction converts into a visible brightness change.
A flat mirror at normal incidence returns linear polarization unchanged. Tilt the surface by α about an axis at 45 degrees to the polarization, and the reflected polarization rotates by 2α; that is pure geometry and holds for any coating. A polarization element spanning the lens pupil keeps the rotation measurable over the entire field, and the analyzer turns it into grey levels:
There is no shear, so nothing is direction-blind: polish marks, machining texture and gentle undulation all render in one shot. That is why wafer backside inspection uses this geometry, with machining-induced undulation and marks appearing as bright areas against a dark field. The trade is edge sharpness: on a 40 nm step the boundary is visible but blurred, because the signal is the slope of the transition, not the step itself. For the finishing side of machined texture, this walk-through of ball-end-mill surface finish on tool steel goes deeper. Note the contrast with cross polarization for surface glare: there the polarizers delete the specular return; here the specular return is the signal.
Which one shows a 40 nm step sharply, and which one shows warp?
Use DIC when the defect is a sharp step or scratch edge, where on a 40 nm reference step it shows a crisp bright-dark boundary, and use pupil division when the defect is gentle tilt, undulation or polish marks across a mirror, where DIC shows only part of the structure.
| Surface feature | DIC image | Pupil-division image | Winner |
|---|---|---|---|
| 40 nm step edge | Crisp dark-bright boundary | Step visible, edge blurred | DIC |
| Scratch 0.2-5 µm wide | Strong line pair, one polarity | Visible as wall tilt | DIC |
| Mirror warp, 0.1 degrees over 10 mm | Partial, directional | Smooth signed shading | Pupil division |
| Polish marks, machining texture | Only across the shear axis | Full texture at once | Pupil division |
| Feature aligned with the shear axis | Invisible until rotated | Still visible | Pupil division |
Decide by defect, then by coverage:
- Wafer frontside scratches, cracks and crystal defects: DIC at 1x and above.
- Wafer backside, large glossy panels, warp and polish texture: pupil division at 0.1x to 0.8x, whole area in a few frames.
- Both defect classes on one part: pupil division to find the zone, DIC to size the defect.
- Scratches with one known direction, such as lap marks: pupil division, or DIC plus a 90 degree second pass.
Interference units of this class are offered from 0.1x to 16.5x optical magnification, pupil division carrying the low end from 0.1x to 4.125x and DIC the high end from 1x to 16.5x, with the sensor format part of the order; 1.1-inch sensor options extend the field further. For how far ordinary lighting stretches on specular wafers before phase methods take over, see reading laser-marked OCR on reflective wafer surfaces.
How large a field of view can you cover?
At 0.1x magnification a pupil-division unit on a 2/3-inch sensor covers 88 x 66 mm, a 110 mm diagonal, while DIC units sit at 1x to 16.5x, so their fields are millimetre scale: 8.8 x 6.6 mm at 1x on the same sensor.
Take a 2/3-inch sensor of 8.8 x 6.6 mm (11 mm diagonal) with 3.45 µm pixels. At 0.1x: field = 8.8 / 0.1 = 88 mm by 6.6 / 0.1 = 66 mm, diagonal 11 / 0.1 = 110 mm, and the object-space pixel is 3.45 / 0.1 = 34.5 µm. At 1x the same sensor gives 8.8 x 6.6 mm at 3.45 µm per pixel. A 1.1-inch sensor at 0.1x pushes the long edge past 150 mm. The deal is explicit: nanometre vertical sensitivity buys you coarse lateral sampling, 34.5 µm per pixel at the wide end.
Stitching covers the rest. Four lanes of 88 mm give 352 mm, enough for a 300 mm wafer with overlap; at 66 mm per row that is five rows, about 20 frames per face. Resolution sets the ceiling in the other direction: resolving a 10 µm scratch width inside an 88 mm frame needs 88 / 0.010 = 8,800 pixels across, which is tiling or line-scan territory. Check your own numbers with the camera resolution calculator and the guide to object space pixel size from magnification.
Can you build this from standard lights and lenses?
No, true DIC needs the shearing prism and bias integrated with the lens, so it comes as a unit; but micron-scale scratches, machining marks and stain that do scatter are reachable today with catalog coaxial, flat-diffuse and dark-field parts.
Draw the line at scatter. If the defects scatter light, meaning machining marks, digs, stains and ragged scratches from a few hundred nanometres upward, standard geometry delivers: coaxial lights for flat shiny fields, telecentric lenses with a coaxial port when magnification must hold constant, low-angle and dark-field sets for raised edges. If the defects are sub-micron relief on a true mirror, steps, polish texture, warp, order an interference unit. A bolt-together replica of DIC does not exist, because the shear and the bias live inside the optics.
For the standard route, the catalog already carries the pieces. The LS-SPC-0116 natural light panel measures 94 x 94 x 17 mm and runs on 24 V, giving a mirror one uniform diffuse field in which every scattering defect appears at once.
- Size
- 50-100 mm
- Voltage
- 24 V
The LS-TCL-2194 is a 0.5x C-mount telecentric lens with a 110 mm working distance, built for 1-inch sensors, with the coaxial light port built in. At 0.5x the object-space pixel is simply double the sensor pixel, so a 3.45 µm camera pixel samples 6.9 µm on the part.
- Sensor format
- 1"
- Mount
- C-mount
- Image circle
- 11-16 mm
- Type
- Object-side telecentric
For compact coaxial units on small flat parts:
- Size
- 50-100 mm
- Voltage
- 24 V
- Size
- 50-100 mm
- Voltage
- 24 V
The LS-CXL-0027 is a 60 x 60 mm coaxial light with 4.4 W red and 7.2 W white channels; the LS-CXL-0036 is a high-precision coaxial light at 74 x 50 x 50 mm. Both families sit in the lights finder beside the lens finder, and the lighting configurator walks the whole standard route. When the job is genuinely nanometer relief, the interference units are special-order: the special lights range includes customized builds, and a quote request carrying your step height, tilt range and field size is the path to a unit from 0.1x to 16.5x.
What camera and light settings do you need?
Use a monochrome camera, a narrowband green source near 525 nm for quantitative DIC, and plan for low throughput: polarizer, beamsplitter and analyzer together pass roughly a third of the light, so budget longer exposure or strobed overdrive.
Colour choice is a decision, not a preference. Narrowband green with a monochrome camera gives one calculable phase scale, so grey level maps to height difference; white light produces colour fringes that a person reads well and software converts poorly. An 8-bit camera covers the 0.09 to 0.91 swing of the 40 nm example with room to spare; pupil division's gentle shading across a warped field rewards 10 to 12 bits. On LED colour trade-offs generally, which LED colour gives the best contrast works through the same logic.
Exposure arithmetic is one multiplication. A part that saturates in 200 µs under plain coaxial needs about 600 µs through the DIC path. At 10 frames per second that is free; at 60 it is not, and overdrive strobing without visible flashes buys the factor back. Re-reference the flat-zone grey level every few hours, because prism bias drifts with temperature and shows up as a slowly tilting background.
When is DIC the wrong answer?
Skip DIC and pupil division on matte or scattering surfaces, since both need a specular return, and skip them when you need absolute height numbers: DIC delivers gradient contrast, not a height map, so a profilometer or 3D sensor is the honest instrument there.
A matte finish, roughness above about 26 nm at green, destroys the coherent specular wavefront and both methods return noise. Textured plastics, anodized and painted surfaces belong to dark field and diffuse dome lighting instead. Strong curvature is the second limit: a specular part that sweeps more than a few degrees of tilt across one field saturates the analyzer, so image it in zones or drop the magnification. Translucent or layered parts add a third: back-surface reflections contribute contrast that is not a surface defect, which is a false call until qualified. And when the deliverable is a height number rather than an edge, measuring surface height difference from a 3D point cloud is the measurement that stands behind the report.
What should you confirm before ordering an interference unit?
Confirm three numbers before you commit: the smallest step you must see, which sets magnification and bias; the steepest tilt in your field, which sets the analyzer offset; and the field size per shot, which sets the stitch count.
- Smallest step, in nm. Anything from about 4 nm up is in range at green; below 10 nm buy the higher magnification.
- Steepest tilt, in degrees. Tilt beyond about twice the analyzer offset saturates the image, so the offset is a spec, not a knob.
- Scratch orientation versus the shear axis, for DIC. Plan the 90 degree second pass in the cycle time, or choose pupil division.
- Field per frame and stitch count. 88 x 66 mm at 0.1x on a 2/3-inch sensor; four lanes by five rows covers a 300 mm wafer face.
- Light colour and camera bit depth. Narrowband green plus monochrome for numbers, white for human review, 10-bit or better for warp shading.
- Fixture flatness. To both methods, a tilted fixture is indistinguishable from a warped part.
- A trial on real defects, not reference samples. Reference steps image beautifully; your polish marks decide the order. And when every unit fails on assembly scratches, the process is the defect: surface protection against assembly scratches is the cheaper fix.
Bring the numbers from items 1 to 3 to a quote request and the unit gets sized correctly the first time. For every defect that scatters, start instead with the lighting configurator; questions the forms do not answer, route to expert support with a sample in hand.
Questions
Can a coaxial light see nanometer scratches if I just increase exposure?
No. A step 40 nm deep on a mirror barely changes the coaxial return, because the surface stays optically smooth; a longer exposure raises the background and the noise together. Coaxial lighting reaches defects that scatter or block reflection, which in practice means scratches from a few hundred nanometres up. Below that, only phase- or slope-sensitive methods, DIC and pupil division, manufacture contrast.
Does DIC work on transparent parts like glass and sapphire?
Yes, in reflection. DIC reads the reflected wavefront, so the part only needs a polished specular surface; sapphire substrates carrying 30 nm laser grooves are a standard demonstration of the technique. One caution on layered or translucent parts: interfaces below the surface also reflect and add contrast that is not a surface defect, so qualify findings with a second pass at 90 degrees to the first shear direction.
What wavelength should the DIC source be?
Green near 525 nm with a monochrome camera gives one calculable phase scale; the 40 nm example above assumes it. White light produces colour fringes that a person reads easily but software converts poorly. Shorter wavelength raises sensitivity in proportion: the same 40 nm step is 55 degrees of phase at 525 nm and about 61 degrees at 470 nm, so blue buys roughly ten percent more swing at the cost of light output.
How do I cover a 300 mm wafer with a 110 mm diagonal field?
Stitch. An 88 x 66 mm frame at 0.1x needs four lanes across a 300 mm wafer, since 4 x 88 = 352 mm with overlap to spare, and five rows along the wafer at 66 mm per pass, so about 20 frames per face; inspecting both faces doubles that. Keep the prism bias setting unchanged across every frame so the grey levels stay comparable in the stitched mosaic.
Is polarized pupil division just crossed polarization with a different name?
No. Crossed polarizers remove the specular return so scatter stands out; the mirror surface itself stays featureless. Pupil division works with the specular return, converting local tilt into brightness through polarization rotation, so the flat mirror is the signal carrier rather than the noise. That is why it renders warp and polish texture that crossed polarizers would show as uniform black.