COG mark point alignment: optics for a 5 µm positioning tolerance

Chip-on-glass (COG) bonding attaches an IC to LCD glass through anisotropic conductive film, and the interconnect survives only when the IC marks and the glass marks line up within a few micrometres. This guide sizes the vision side of such an alignment station for a ±5 µm positioning tolerance: it turns the tolerance into a pixel budget, the budget into magnification, and the magnification into a camera, lens and light choice. The worked example pairs a 4000 × 3000, 1/1.7" sensor with a 1.36x object-side telecentric lens, and every check is arithmetic you can repeat with your own mark spacing and stage specs.

The decision, in numbers

For this alignment, size the optics so one pixel is 1-1.7 µm in object space: between one third and one tenth of the tolerance, with about 1.3 µm as the working target. On a 4000 × 3000, 1/1.7" sensor that calls for at least 1.1x magnification. The fit is a 1.36x object-side telecentric lens at 73 mm working distance: a 5.3 × 4.0 mm field, 1.3 µm per pixel, 3.8 pixels across the tolerance. At that pixel size the vision chain contributes about 0.5 µm at 3σ; hold stage repeatability to 3 µm and datum the glass height, and the full budget closes near 3.2 µm RSS. Manual alignment through a loupe stops around 20-50 µm and drifts with fatigue, which is why COG attach is vision-only. The sensor side of that budget comes from the area scan cameras range.

How a tolerance becomes a pixel budget

A geometric pattern matcher locates a taught mark to roughly one pixel. A grayscale centroid or edge fit on the same mark repeats to 1/10 pixel when the edge is clean and the light steady. Position accuracy is pixel size times matching repeatability, so the pixel size is the design variable. Machine vision sizes it by a fixed rule: keep the object-space pixel between one third of the tolerance (alignment work) and one tenth (measurement work). The same arithmetic that asks 10 µm pixels for a ±0.1 mm measurement asks 1.7 µm pixels or better here.

p is the object-space pixel size, W the sensor width, m the magnification, N the horizontal pixel count and T the positioning tolerance. With W = 7.2 mm, N = 4000 and T = 0.005 mm: m ≥ 3 × 7.2 mm / (0.005 mm × 4000 px) = 1.08, so 1.1x is the floor. The one-tenth rule would demand 3.6x and shrink the field to 2.0 mm; that is measurement grade, not alignment grade — size to one third and let sub-pixel matching carry the rest. The derivation is set out step by step in calculate object-space pixel size, and the general sizing logic in sizing camera resolution by the pixel rule.

Worked example: 1.36x on a 4000 × 3000 sensor

Apply m = 1.36x to the sensor: field of view = 7.2 mm / 1.36 = 5.3 mm wide and 5.4 mm / 1.36 = 4.0 mm tall. Object pixel = 5.3 mm / 4000 px = 1.3 µm. The checks:

• 1.3 µm against the 1.7 µm budget: pass.
• the tolerance spans 0.005 mm / 1.3 µm ≈ 3.8 pixels, against the 3-pixel floor: about 25% headroom.
• Matching at 1/10 pixel adds 0.13 µm of algorithm error; a conservative 1/3 pixel is 0.44 µm. Both sit far below tolerance.
• Rotation: two marks 4 mm apart, each located to 0.5 µm, give θ = 0.5 µm / 4 mm = 0.000125 rad ≈ 0.007°. Position, not angle, is the binding constraint.

Why telecentric, not a fast macro

Between parts, the glass and film stack height varies; take ±0.1 mm as a typical spread. An entocentric lens focused at 100 mm changes scale by 0.1 / 100 = 0.1% over that height swing, and across a 2 mm mark-to-mark span 0.1% is 2 µm of false position — 40% of the entire budget before any other error contributes. An object-side telecentric holds magnification constant across its telecentric depth, so part height becomes a defocus question only. The trade is laid out in entocentric vs telecentric lenses.

Depth of field then sets the f-number. The estimate t = 2 × F × p × (1 + m) / m, with p the object pixel: at F5.6, 2 × 5.6 × 1.32 µm × 2.36 / 1.36 ≈ 26 µm total, or ±13 µm; at F8, about 37 µm, or ±18 µm. Datum the glass on a hard stop so the mark plane repeats within ±10 µm and the mark stays in focus at F5.6-F8. The f-number trade-offs continue in working f-number: light, DOF and resolution.

Diffraction sets the honest floor: blur diameter ≈ 2.44 × 0.55 µm × F × 2.36 / 1.36 gives 9 µm at F4 and 19 µm at F8 in object space — several pixels wide at this magnification. Symmetric blur does not shift a centroid, so located position survives it; what blur costs is edge sharpness for the matcher score. Keep the marks high-contrast and prove repeatability on static captures before bonding.

Both lens candidates below cover the 1/1.7" sensor on a C-mount. At 1x and 108 mm the field opens to 7.2 mm and the pixel grows to 1.8 µm: 2.8 pixels across the tolerance, under the 3-pixel floor, which fits that combination to ±10 µm class jobs where the budget allows 3.3 µm pixels. For the tighter class, take the 1.36x. If the fixture needs more clearance than 73 mm, the telecentric lenses range includes longer-working-distance magnifications to re-run the formula against.

All 2 share: Mount C-mount. Values as printed in each product's data sheet; open a part for its drawing and full specifications.

The LS-TCL-0811 is the lens this example sized: object-side telecentric, 1.36x, 73 mm working distance, C-mount, 2/3" image circle, with a coaxial light port.

Camera, light and exposure

Capture is one still frame per cycle and nothing moves during it, so a rolling shutter is fine and frame rate is nearly free: the USB3 camera here captures at 30 fps and the GigE twin at 9 fps, both far above the one frame every second or two that the loop needs. Choose the interface by the host, not by speed. Mark location prefers a mono sensor and a hard edge; color adds nothing on a chrome mark.

The lens's coaxial port feeds LED light down the optical axis. Chrome marks reflect it straight back bright while plain glass returns almost nothing, which is exactly the contrast the matcher wants. Drive the LED from a constant-current controller: brightness drift moves match thresholds and, with them, located edges. Re-teach the marks after any lamp change. Exposures in the single-digit milliseconds are normally sufficient through a coaxial port; set the histogram so mark edges use most of the range without clipping.

Either camera runs this example — the GigE at 9 fps, the USB3 at 30 fps.

Closing the loop: calibration and error budget

Calibrate each camera to the stage — a calibrated glass plate, or the stage's own commanded moves — and a telecentric's residual distortion lands at the 1-2 µm level across this field. Where slow drift comes from later, and how to catch it, is the subject of micron-level image measurement accuracy.

The per-axis budget:

RSS = √(0.5² + 1.0² + 3.0² + 0.5²) = 3.2 µm; the worst-case sum is 5.0 µm. The thermal line has a basis: aluminium grows 23 ppm/°C, so a 100 mm stage path moves 2.3 µm per °C, and ±0.2 °C of control keeps that term near 0.5 µm.

Then the loop: teach the nominal mark positions, match each part, convert the offsets to ΔX, ΔY and Δθ on the stage, command the correction, and repeat until the readback sits at or under 2 µm before bonding. Set the matcher's score gates and accept thresholds deliberately — the classic failure modes are catalogued in pattern-match tolerance settings. Accept the station on two tests: 50 static captures with 3σ at or under 1 µm, and commanded stage steps equal to the tolerance that read back within 1 µm.

Parts list and next steps

• Camera: LS-CMR-0166 — 12 MP, 4000 × 3000, 30 fps, USB3 — or LS-CMR-0053 — 12 MP, 9 fps, GigE, C-mount.
• Lens: LS-TCL-0811 — 1.36x object-side telecentric, 73 mm working distance, coaxial port.
• Compare machine vision lenses in the lens finder, and sensor options in the camera finder.

If your marks, spacing or tolerance differ, re-run the section-two formula with your numbers; it takes a minute and changes the magnification, not the method. To turn a different set of numbers into a parts list, use the contact page, and the neighbouring guides live under the systems topic.