The Error Frame Temperature Leaves in Two-Sided Chromatic Confocal Thickness Measurement

OPTICS

A 1 K change in the steel frame between two heads moves the thickness reading by 1.2 μm, leaving little room for catching a 2 μm deviation.

Taking a 2 μm thickness deviation in transparent covers and coatings, I calculated how chromatic confocal works, how two-sided two-channel measurement removes the refractive index, and what heat and tilt take away, along with spot offset.

About 12 min read

An injection-molded lens dome or a glass cover is a transparent sample with two smooth faces, so a little light reflects from the surface like a mirror and most passes inside. One face of 1.5-index glass returns only 4 % of the light at normal incidence, by Fresnel reflection¹. On parts like these, thickness separates good from bad more than scratches do. When a coating or adhesive layer comes out 2 μm thinner than designed, color and shape stay the same, and the defect surfaces late as a shifted focus or weak bonding.

A 2D camera has no clue to grab this deviation with, because the only thing that changed is distance along the optical axis. That makes a chromatic confocal² sensor, which measures distance directly, the candidate, and products have recently appeared that tie two heads to one controller to measure the top and bottom of a sample together. I weigh this setup against a single number, 2 μm: whether the principle can handle it, and what two-sided measurement removes and what it brings in.

Why a camera cannot see 2 μm of thickness

What a camera records is the brightness distribution of reflected light. If thickness changes by 2 μm while the surface keeps its color and pattern, the brightness distribution stays the same and the image keeps no trace of the change. Focus cannot read it either. Estimating the depth of field³ of an inspection lens with a numerical aperture of 0.05 at a 0.55 μm wavelength gives 0.55 ÷ 0.05² = 220 μm, so a 2 μm difference is under 1 % of that depth.

Indirect judgments from shadows or interference fringes exist, but the results swing with lighting angle and surface condition. When thickness itself is the quality measure, I do not drag out the hunt for clues in images and move on to a sensor that measures distance along the optical axis directly. Knowing how that sensor measures is what decides how far the 2 μm reading can be trusted.

Wavelength measures the distance

A chromatic confocal sensor uses a lens built with deliberately large axial chromatic aberration⁴. Send white light through it and each wavelength comes to focus at a different distance, and only the wavelength focused exactly on the target surface passes back through the pinhole to the spectrometer⁵. Converting the position of the strongest wavelength in the spectrometer into distance completes the measurement. Mitutoyo’s US patent US10197382B2 (priority date October 20, 2016; granted February 5, 2019) covers a sensor that focuses light of different wavelengths at different distances through an objective lens and calculates the target position from the light returning from the focused spot, with an added observation unit for viewing the measurement spot.

In a transparent sample, the top and bottom faces each return a different wavelength, so two peaks appear in the spectrum and the gap between them becomes the thickness. That gap is an apparent value bent by refraction inside the sample, though, close to the real thickness divided by the refractive index⁶ in the paraxial approximation. When a 1.5-index coating thins by 2 μm, the apparent gap shrinks by only 1.33 μm, and entering the index 0.01 off for a 100 μm coating throws the thickness off by 0.67 μm. That 0.67 μm is no small number.

Weak reflection between layers is another problem. An interface where indices of 1.5 and 1.6 meet returns only 0.1 % by Fresnel reflection, a peak about 38 times weaker than the 4 % at an air-glass boundary. Peaks like that easily blend into noise, so the controller needs both a minimum peak intensity and a limit on how many peaks to look for. Whether peaks separate reliably in a multilayer sample, with index differences and surface roughness varying by layer, cannot be guaranteed before a sample test.

Measure both sides and the refractive index drops out

Placing two heads facing each other above and below the sample changes the calculation. The top head measures the distance to the top face, the bottom head the distance to the bottom face, and subtracting both from the gap between the heads gives the thickness. Light never passes through the sample, so the refractive index drops out of the calculation. Being able to measure opaque samples the same way is a use the multi-peak method lacks.

Micro-Epsilon has released the two-channel controllers IFC2412 and IFC2417, and Instrumentation Monthly reported on February 23, 2026 that they enable two-sided thickness measurement with a single controller. According to the company, the IFC2412 offers a measuring rate adjustable up to 8 kHz with resolution from 2 nm, and the IFC2417 measures up to 25 kHz and handles up to five transparent layers by multi-peak measurement. In contrast to setups that matched the values of two controllers in software, both channels are tied together by calculation functions inside one unit.

As the figure shows, the two heads must face each other on the same axis for the thickness to be the thickness at one point. The structure removes the refractive index but brings in new variables: the gap between the heads and the alignment of the two spots. In a setup handling 2 μm, these two are the first to waver.

The frame between heads moves 1.2 μm per kelvin

Thickness is the gap between the heads minus the two distances, so when the gap changes, the thickness changes with it. Assuming the steel frame joining the heads is 100 mm long with a linear expansion coefficient⁷ of 11.7 × 10⁻⁶/K, a 1 K rise stretches the gap by 1.17 μm and the thickness reading drops by 1.17 μm though the sample is unchanged. Holding measurement error to a tenth of a 2 μm deviation, 0.2 μm, works out to keeping the frame within 0.17 K. With an aluminum frame, whose expansion coefficient is about twice that of steel, the 1.17 μm per kelvin nearly doubles.

Holding 0.17 K continuously on the floor is hard. So I choose to put a reference sample⁸ of known thickness between the heads at set intervals and re-zero the gap value. The interval only has to keep the temperature change between reference checks under 0.17 K, so a log of frame temperature lets that interval be calculated too. Building the frame from a low-expansion material is an option, but the reference check is still needed then.

What tilt and spot offset take

When the sample tilts relative to the measuring axis, the length the two heads measure becomes the thickness divided by the cosine. A 1.0 mm cover tilted 1° gives a cosine error⁹ of 0.15 μm, 0.61 μm at 2° and 1.37 μm at 3°. A 3° tilt alone eats two thirds of a 2 μm margin, so the parallelism of the fixture holding the sample becomes part of the measurement precision. The fixture is part of the instrument.

Sideways offset between the two spots¹⁰ is a problem of the same size. With the top and bottom spots 50 μm apart and the bottom face sloping 2 %, the heads measure the thickness at different places, an error of 50 μm × 0.02 = 1 μm. On curved faces like a lens dome, the slope grows away from the center and so does this error. I would check spot offset with a wedge sample of known slope alongside the reference sample, and look first at heads with an observation unit for aligning the measurement spot by eye.

Speed goes on the same table. If the sample passes at 100 mm per second while a line is scanned, 8 kHz yields a point every 12.5 μm and 25 kHz every 4 μm. Raising the measuring rate packs the points closer, but the light time per measurement shrinks, so it is worth watching whether the peak weakens on a transparent face that reflects only 4 %.

A two-sided setup for 2 μm deviations, and the line with structured-light 3D

Before building the two-sided gauge, put three things in one table: the size of the deviation to judge, the conditions of the two heads, frame and fixture, and the controller parameters. Head WD¹¹ and measuring range differ by model and have to be checked on the datasheet, and in a facing-head setup the gap must leave room for loading and unloading the sample and for the reference sample as well. With a 0.2 μm error budget for judging 2 μm, how heat and tilt split that budget with spot offset is the center of the table.

Item	① Minimum defect size	② Optical setup · lighting · WD	③ Algorithm parameters
Thickness deviation	2 μm deviation in coating or cover thickness (0.2 μm error budget)	Two chromatic confocal heads facing each other on one two-channel controller. Check head WD and measuring range on each model’s datasheet; keep room in the gap for loading and for the reference sample	Thickness = head gap − top distance − bottom distance. No refractive index needed (the multi-peak method errs 0.67 μm on a 100 μm coating for a 0.01 index error)
Heat · frame	1.17 μm per kelvin (100 mm of steel assumed)	Log frame temperature; consider low-expansion material	Set the reference-sample interval so frame temperature drifts less than 0.17 K between checks
Tilt · spots	1.37 μm at 3° tilt; 50 μm spot offset × 2 % slope = 1 μm	Control fixture parallelism; check spot alignment with a viewing unit	Set a minimum peak intensity and a peak-count limit (4 % reflection at a glass face, 0.1 % between layers). At 8 kHz and 100 mm/s, a point every 12.5 μm

With all three rows filled, it becomes possible to count whether a 2 μm deviation is judged within a 0.2 μm error budget. Chromatic confocal resolution may sit at the nm level, but heat and tilt eat the margin at the μm level, and so does spot alignment.

There are also places where structured-light 3D fits better. When the shape of a whole surface must be seen at once, as with solder bumps or molded surfaces, structured-light 3D¹², which reconstructs the surface in one go, is faster than chromatic confocal measuring point by point. Chromatic confocal is strong at repeatedly measuring the thickness of a few set points at high resolution. Which side wins on a real part depends on surface reflectance and the number of points needed, so it cannot be guaranteed before a sample test.

Breaking thickness error down by cause, from heat to tilt to alignment, is the starting number for an engineer setting inspection specs for transparent parts, and in precision measurement the same breakdown becomes the backbone of a calibration plan. When choosing the gauge, get a frame temperature plan and a reference-sample routine along with the sensor resolution figure, or that resolution will not survive on the floor. This article explains technology and is not investment advice. See our Disclaimer.

Field Note

If a coating coming out 2 μm thinner than designed had to be caught, my first step would be preparing a reference sample of known thickness. Putting it between the two heads and logging the thickness every hour for a day beside the frame temperature turns the microns per kelvin into a measured value. Next, a wedge sample would show the offset between the two spots, and tilting the fixture 1°, 2° and 3° would confirm that the cosine error follows the calculation.

Field Checkpoints

Once a day of reference-sample records and the tilt test results are in, whether two-channel, two-sided measurement goes into production comes down to the questions below.

• WD: do loading space and the reference-sample position fit inside the gap between heads while keeping each head’s WD?
• Is frame temperature logged, with the reference-sample interval set to keep temperature change under 0.17 K? (1.17 μm per kelvin for 100 mm of steel)
• Has fixture tilt been measured to keep the cosine error within budget? (1.37 μm at 3° for 1.0 mm thickness)
• Has the offset between the two spots been checked with a wedge sample? (50 μm × 2 % slope = 1 μm)
• If the multi-peak method is used, has the basis for each layer’s refractive index been confirmed?
• Is the minimum peak intensity set for transparent-face reflection (4 %) and interlayer reflection (0.1 %)?
• Does peak intensity stay above the minimum after the measuring rate is raised?

Frequently Asked Questions

Should I choose single-head multi-peak or two-channel two-sided measurement?

For a coating on an opaque substrate whose bottom face light cannot reach, measuring from one side by multi-peak is the only way, and the layer indices must then be accurate. If heads can sit above and below the sample and the index is uncertain, two-channel two-sided measurement removes the index problem. In exchange comes the burden of managing frame expansion and spot alignment.

With 2 nm resolution, isn’t 2 μm easy to measure?

Resolution is the smallest change the sensor can tell apart, while the accuracy of a thickness value is set jointly by the frame and fixture together with alignment. As calculated above, 1 K of frame change can make 1.17 μm and a 3° tilt 1.37 μm, so my criterion is whether the sum of these errors stays within 0.2 μm.

Is a higher measuring rate always better?

Points get denser, but the light time per measurement shrinks. On a transparent face reflecting only 4 % or an interface reflecting 0.1 %, the peak can weaken, so it is steadier to calculate the point spacing needed first and start at the lowest rate that achieves it.

Glossary

• [1] Fresnel reflection Fresnel reflection
  The return of part of the light at a boundary between materials of different refractive index. At normal incidence it is ((n₂ − n₁) ÷ (n₂ + n₁))². ↩
• [2] Chromatic confocal chromatic confocal
  Measuring distance to a target with a lens whose focal distance differs by wavelength, a pinhole and a spectrometer. ↩
• [3] Depth of field DOF
  The range of distance that still looks in focus. For a diffraction-limited lens it is roughly the wavelength divided by NA squared. ↩
• [4] Axial chromatic aberration longitudinal chromatic aberration
  Different wavelengths coming to focus at different points along the optical axis. Chromatic confocal enlarges it on purpose. ↩
• [5] Spectrometer spectrometer
  A device that splits light by wavelength and measures the intensity of each. ↩
• [6] Refractive index refractive index
  How much slower light travels in a material than in vacuum. Glass is around 1.5. ↩
• [7] Linear expansion coefficient CTE
  The fractional change in length per kelvin. The text assumes 11.7 × 10⁻⁶/K for steel. ↩
• [8] Reference sample master gauge
  A sample of precisely known thickness, measured at set intervals to re-zero the measuring system. ↩
• [9] Cosine error cosine error
  The error from tilt between measuring axis and target, which lengthens the measured value by a factor of 1/cos. ↩
• [10] Measurement spot measuring spot
  The small point of light on the target where the sensor actually measures distance. ↩
• [11] WD working distance
  The distance from the front of the head to the target surface. It also decides room for loading and for the reference sample. ↩
• [12] Structured-light 3D fringe projection
  Projecting a pattern and reconstructing the 3D shape of a whole surface at once from how the pattern bends. ↩

References

• Micro-Epsilon, “New possibilities in confocal measurement technology” newsroom article (IFC2412 at 8 kHz and from 2 nm, IFC2417 at 25 kHz with five transparent layers, two-sided thickness with one controller). micro-epsilon.com
• Instrumentation Monthly, “New compact two-channel confocal controllers enable two-sided thickness measurements using just one controller”, 2026-02-23. instrumentation.co.uk
• Mitutoyo Corp., US10197382B2 “Chromatic confocal sensor”, priority date 2016-10-20, granted 2019-02-05. Google Patents

Related reading — How 3D Structured Light Measures a Solder Lead Lift That Brightness Cannot Show · A Depth Camera’s Error Grows with the Square of Distance · How Liquid Lens Autofocus Solves Focus on High-Mix, Low-Volume Lines