Reading laser-marked OCR on reflective wafer surfaces

Laser marks sit only microns above a wafer polished to a mirror finish, and that finish is what kills ordinary OCR lighting: the polished field returns the light straight into the lens and saturates the image, while the shallow, matte mark scatters almost nothing back. This how-to gives the fix in the order that matters. First choose the lighting geometry that separates the mark's diffuse reflection from the wafer's specular one. Then size the pixels and the exposure with arithmetic you can check. Last, prove readability with measured grey levels rather than your eye. It is written for process, test and traceability engineers adding character OCR to wafer handling, on a new station or retrofitted into an existing tool.

Why laser marks vanish on a polished wafer

The wafer and the mark obey two different reflection laws. Polished silicon is specular: an incoming ray leaves at the mirror angle, which is the behaviour behind bright field versus dark field lighting. The laser mark is a re-melted, micro-rough surface that scatters light diffusely in all directions. The camera sees only the light that reaches its lens, so geometry alone decides which of the two surfaces sends light there.

Light a wafer from a shallow angle and part of the field's specular return always finds the aperture. The background rides near saturation while the mark, which scatters weakly toward the camera, sits a few grey levels above it. The same code that reads easily on a matte test coupon then fails on the production wafer. It is the identical trap met when reading QR codes on reflective and curved surfaces, and the way out is the same: control the specular return instead of fighting it in software.

Three geometries do that controlling. Put the light on the optical axis, take it to a glancing angle, or filter the specular part with polarization. Exposure, gain and thresholds come afterwards, and none of them rescues the wrong geometry.

Choose the geometry by mark depth

Coaxial lights are the first choice for shallow, flat wafer marks. Light travels down the camera's own axis through a beam splitter, so the mirror-flat field returns as an even, controlled bright background and the matte mark interrupts it as dark characters. A typical 50:50 splitter passes half the light on the way down and half of that again on the way back, so about a quarter of the emitted light reaches the sensor; budget current or strobe time for it.

Deep black engravings behave the opposite way: their walls catch glancing light and shine, while the polished field skips it away into darkness. Light them with low-angle bars from one or four sides, exactly as in DPM data matrix lighting on metal or low-angle lighting for dents and pits. The characters appear bright on black, with slightly widened strokes because the engraving walls have slope.

Bow, warp or a chuck that is not perfectly level moves the specular direction across the field, and a coaxial unit then leaves bright and dark bands. A dome or another very diffuse source returns the field bright from every direction and restores an even background, as covered in dome lights for shiny, curved and wrinkled parts.

Whatever specular return survives the geometry, polarization removes it. A polarizing sheet on the light and a crossed analyzer on the lens block the specular component, which keeps its polarization, while passing the depolarized diffuse light of the mark. The same principle drives crossed polarized lighting for transparent film defects and glare-free imaging of shiny metal pins. Each sheet passes well under half the light, so add exposure to compensate.

Coaxial lights sized from a 60 x 60 mm mark field up to a 140 x 140 mm wider view:

The 60 x 60 mm unit above prints 4.4 W in red and 7.2 W in white. Run white while proving the geometry, then switch to red for production on a mono sensor, following the contrast logic in light color and monochrome camera contrast. Multi-angle and zone-sequential variants for edge marks and ring codes sit in special lights.

Worked example: 1 mm marks across a 150 mm wafer

The case: a 150 mm wafer read in one 160 mm square field, with 1 mm laser-marked characters, on a stage moving at 50 mm/s. Substitute your own wafer size, field and stage speed; the method is identical.

Inputs, and where each comes from: the 160 mm field buys 5 mm of alignment margin around the wafer; the sensor delivers 4000 × 3000 pixels; the legibility rule of 20 to 30 pixels of character height is the range in which OCR engines hold reliably, with strokes merging below about 20 pixels.

The rules in words: divide field width by sensor width to get the scale; divide character height by the scale to get pixels per character; divide the one-pixel blur budget by pixels-per-millimetre times stage speed to get the longest exposure.

Result: 25 pixels of character height, inside the legibility range but near its floor, and 0.8 ms of exposure, which needs a strobed or heavily driven light. Three decisions follow:

• 1 mm characters, full field: readable, with the margin coming from lighting geometry rather than pixel count.
• 0.5 mm characters: at the same scale they reach only 12.5 pixels, below the floor. Tighten the field to 4000 px / 50 px/mm = 80 mm and read the scribe region alone, or step up the sensor.
• 0.8 ms not available from your light: read stationary instead. Exposure then stretches into the tens of milliseconds, and light current, not timing, sets the brightness.

Re-run the first two lines with your own wafer and character sizes here:

One camera covers the whole example field in a single shot:

A mark-only field shrinks to 10 or 20 mm and the resolution bill shrinks with it. A 6 MP GigE mono camera with a 1/1.8" sensor then leaves headroom to spare:

Compare the rest of the range in the area scan cameras category or the camera finder.

Set up the station and prove it reads

1. Fix the geometry before anything else. Match the table above to your mark depth. Mount a coaxial light with its splitter axis exactly on the lens axis; a degree of tilt smears the bright field across the wafer.
2. Choose mono and one wavelength. Mark against field is a brightness difference, so a mono sensor puts every photosite on it; keep color cameras for tint sorting. Run set-up in white, production in red or whichever wavelength your coating favors.
3. Compute scale and exposure from your own numbers. Apply the three formulas above with your field, sensor and stage speed, and set the blur budget at one pixel; two pixels visibly softens stroke edges.
4. Set exposure before gain. Take the shortest exposure that keeps the field comfortably below saturation, and add gain only after geometry and exposure are settled, because noise eats stroke contrast first.
5. Measure contrast numerically. Sample two patches of at least 100 x 100 pixels, one inside a stroke and one on the field beside it. Require a mean difference of 25 grey levels or more at production exposure, and log both means with every image. The patch-mean habit also appears in measuring wafer defect area without a background.
6. Pair strobe with the right shutter. For on-the-fly reads, fire the light inside the sub-millisecond exposure and use a global shutter camera; a rolling shutter smears a short flash across rows. The timing rules of overdrive, duty cycle and flash window are set out in strobe timing for machine vision lights.
7. Prove it holds before release. Rotate the wafer a few degrees each way, run wafers from two mark batches, and re-measure the patch contrast at each step. Contrast that survives those variations is what separates a production read from a demonstration, the argument made in why adequate contrast is not enough.

When the usual answer is wrong

• Reading near the edge or bevel. Curvature swings the specular return out of the coaxial geometry; only a dome that wraps the edge restores an even background there.
• Marks under resist or a transparent film. The film adds a second specular layer. Crossed polarizers strip the part that keeps its polarization; birefringent films leak the rest, more with tilt. Test at the steepest angle the station allows before committing.
• A coating that fluoresces. Violet or UV excitation near 365-405 nm with a matched emission filter can light the mark region against a dark background, the approach described in fluorescence inspection lighting.
• You can still change the mark. Specify mark depth and anneal contrast upstream. A slightly deeper engraving buys more grey-level separation than any camera upgrade, and it costs nothing at the vision station.

Checks before ordering

• Mark samples in hand, best and worst from two anneal batches; contrast is decided at the laser as much as at the camera.
• Geometry chosen from the table, not from habit: shallow means coaxial, deep means dark field.
• Pixel scale computed: character height divided by scale gives 25 pixels or more.
• Exposure budget held: one pixel of smear at your stage speed, or a stationary read.
• Light size covers the field at the working distance, and the mount envelope fits the tool.
• Shutter matches motion: global shutter for strobed reads, rolling shutter only with continuous light.
• Strobe-capable controller in the bill of materials if the exposure lands below a millisecond.

Turn the numbers into an order

Once the patch contrast holds on your own wafers, the lighting configurator and the lights finder turn the geometry into an orderable set of parts. If you would rather send mark samples first, start through the contact page.