Glass Core TGV Inspection: Surface AOI vs Subsurface Crack Detection

Evidence disclosure: this page is an independent engineering guide informed by public product information published by Spirox Corporation. It is not a customer case, an installed-base claim, an endorsement or a verified compatibility statement. The camera and lens models actually installed on the referenced equipment, and any purchasing relationship with POMEAS, are UNKNOWN.

A glass core substrate asks three different inspection questions in a single panel. Through-glass vias have to be measured for diameter, position, pitch and taper. The build-up on top of the glass, whether ABF or a photo-dielectric, has to be measured for line width, dielectric openings and pad position, and inspected for surface defects. And the glass itself has to be checked for cracks that never reach either surface. Each question is answerable, but none of them is answered by the same optical configuration, and that is where procurement tends to go wrong: a single station is specified for "glass core inspection" and then has to deliver three results that belong to three different physical measurements.

This page separates the three jobs on physical grounds rather than on marketing grounds, then does the arithmetic that decides which POMEAS lens could be evaluated for the surface part of the work. It states explicitly where a visible-light optical station stops, because the most expensive mistake in this application is a station that images a buried crack clearly enough to look convincing and not clearly enough to be repeatable.

Three jobs on one panel, measured by three different physics

Via geometry is a measurement of a through feature. With parallel backlight behind the panel and a telecentric lens in front of it, the bright disc in the image is the projection of the hole along the optical axis, and its edge is defined by the substrate rather than by the illumination. Diameter, position relative to a datum, pitch and array regularity are all edge-based measurements in a single plane, which is the strongest case for a calibrated telecentric field: image scale does not change when the panel sits a few tenths of a millimetre higher or lower.

Surface and build-up dimensions are a measurement of a reflecting plane. Here the light returns from the top surface, and the contrast comes from the difference between copper, dielectric and glass. The same telecentric lens can be used, usually with coaxial illumination for metal and angled illumination for dielectric edges, but the magnification is an order of magnitude higher than for a whole-panel via array, and the depth of field collapses accordingly.

Subsurface cracks are not a measurement at all in the same sense. They are a detection problem inside a transparent volume: the signal is light scattered from a crack face that is buried tens to hundreds of micrometres below the surface, superimposed on the surface texture that returns far more light. That is a different illumination geometry, a different wavelength band and a different image formation model, and it is the reason the two functions are sold as separate instruments rather than as options on one.

Why a surface station cannot be re-tuned into a crack inspector

A visible-light telecentric lens forms an image of one object plane. Everything outside that plane is defocused and, in a transparent sample, still present in the image as a blurred overlay. A crack under 300 µm of glass therefore contributes a low-contrast, blurred feature sitting on top of a sharp surface pattern, and no change of lens, aperture or pixel count recovers the axial information that would separate them. Depth discrimination has to come from the illumination and detection scheme: illumination delivered through the volume rather than onto the surface, collection restricted to a defined scattering angle, or a confocal or coherence gate that rejects light from outside a thin axial slice.

The published SP7000G specification shows exactly this logic. The system is described as laser tomography using 1200-1800 nm illumination, which is a band chosen for how it propagates in glass and how it interacts with the buried interfaces rather than for surface appearance. It uses four scan heads with built-in 4X objectives and covers a maximum panel of 510 x 515 mm. None of that is a lens specification that a surface AOI station could match by buying a longer working distance or a higher numerical aperture.

What the equipment disclosure establishes, and what it does not

The published table pairs a thickness of scan layer with a minimum detectable crack, and the pairing is the useful part: detection capability degrades as the inspected volume gets thicker. At 250 µm the stated minimum detectable crack is 3 µm, at 500 µm it is 7 µm, and at 800 µm it is 15 µm.

Three cautions apply to that table. First, these are equipment-specific detection claims made under defined settings, not a general physical law, and they cannot be extended to another instrument or another material stack. Second, a minimum detectable crack is a detection threshold, not a dimensional accuracy: it says a crack of that size can be found under those conditions, not that a crack's width can be measured to that value. Third, the single-layer scan time of ten minutes with four heads is a throughput statement about one layer of one panel size, so inspection time has to be recalculated for the layer count, the panel format and the sampling density of any particular line.

Field arithmetic: which magnification fits which part of the panel

For the surface half of the work the constraint is field against resolution, and POMEAS telecentric lenses bracket the range. The published values below are object-side resolution figures for the lens, independent of the camera, and the field of view column is derived by dividing the sensor format by the magnification.

Read the first row and the third carefully together, because they explain most of the confusion in this application. The 0.061X lens covers a 144 x 108 mm field at a 398 mm working distance, which is what makes it usable over a supported panel with a carrier underneath; its object resolution of 87.99 µm means a 40 µm feature is below what the optics resolve, so the station sees the via array as an array and each via as a disc, not as a feature with a measurable edge quality. The 1.0X lens at 85 mm working distance resolves 8.39 µm and covers 8.8 x 6.6 mm, which is a die-level field, not a panel-level one.

Both rows also carry the same warning: pixel scale is not resolution. At 0.061X with a 2448-pixel sensor the arithmetic gives 58.9 µm per pixel, which is finer than the lens can resolve, so the sampling is honest but the optical limit is reached first. At 1.0X with the same sensor the pixel scale is 3.6 µm against an 8.39 µm lens resolution, so the lens is the limit and a higher-resolution camera buys nothing on that lens.

How a tapered via reads under telecentric backlight

A via that is wider at the top than at the bottom does not produce the top diameter in a backlit image. With collimated illumination and a telecentric lens, the silhouette is the narrowest cross-section along the optical axis, because every ray that is blocked anywhere in the hole fails to reach the sensor. The bright disc is therefore the throat of the via, which is usually the dimension that matters for fill and resistance, but is not the entry diameter that a design drawing often quotes.

Measuring both ends means measuring at both surfaces. Focus at the top surface with reflected illumination to find the entry diameter, focus at the bottom of the glass to find the exit diameter, and subtract: the arithmetic is straightforward once both diameters exist. For a hole 100 µm at the top and 70 µm at the bottom through 500 µm of glass, the half-angle is the arctangent of 15/500, or about 1.72 degrees (derived). Doing that on one station requires either a focus axis with a repeatable travel or two stations, and the choice depends on how tight the taper tolerance is: the axial travel needed to resolve the two planes is small compared with the depth of field of the 1.0X lens, so a single-focus solution is usually the right one for taper and the wrong one for throughput.

Lighting and holding a glass panel

Glass is transparent, so the lighting choices do not overlap with those of a metal part. Through-features want parallel backlight, which produces a hard silhouette and no surface texture in the measurement. Surface features want coaxial illumination, because the return from a coated or metallised layer is specular and a ring light alone tends to place the highlight rather than the edge. Scratches and surface-opening cracks want low-angle or dark-field illumination, where the defect scatters into the lens and the flat surface does not. Mixing these on one station means switchable illumination rather than a fixed ring, and it means the recipe has to record which illumination was used for which measurement, because the numbers are not comparable across modes.

A panel that is 510 x 515 mm and under a millimetre thick is also a mechanical problem. It needs support over its whole area, no clamp over the inspection region, and a reference surface flat enough that the part does not tilt out of the depth of field. The published depth of field of the 1.0X lens is 1.00 mm and of the 0.16X lens 18.75 mm, which is generous at low magnification and tight at high magnification; a carrier that is flat to a few tenths of a millimetre is adequate for the former and marginal for the latter. Transparent parts also produce ghost edges from the second surface, which is a well-documented failure mode and worth checking on the actual panel before the recipe is frozen.

A sampling calculation, and what it does not prove

Take a 20 mm horizontal field on a 2448-pixel sensor. Sampling is 20,000 divided by 2448, or 8.17 µm per pixel (derived). A 40 µm surface mark then spans about 4.9 pixels across its width, before blur, contrast and orientation are considered. That number is a reasonable shortlist criterion: a feature spanning five pixels in a well-lit, high-contrast image is a comfortable detection target, and one spanning a pixel and a half is not, whatever the datasheet says about resolution.

What the calculation does not do is establish a probability of detection. Detection depends on the contrast the defect produces against its background, on whether its orientation puts its edge along or across the pixel grid, on the illumination mode and on the decision rule in the software. Those are measured on real panels with known defects, or they are not known at all.

Validation protocol for a surface inspection station

Use panels from the actual process steps, not polished coupons: post-CMP glass, post-ABF build-up and post-RDL panels behave differently under the same illumination. Record surface defect classes separately from subsurface classes, because a single false-positive count that mixes a scratch with a buried crack tells you nothing about either. Measure repeatability across at least nine positions on the panel, including the four corners, and repeat the run at the start and end of a shift to expose calibration drift. Keep a golden panel with defects confirmed by cross-section, and re-image it whenever the recipe, the illumination or the software revision changes. Before any performance claim is published or quoted, capture the sample images and a signed acceptance table.

Where the boundary sits

POMEAS telecentric lenses and zoom optics can be evaluated for the visible-light, surface-side part of this work: via diameter and position under backlight, build-up dimensions under coaxial or angled illumination, and edge geometry on the panel outline. They are not presented as a substitute for near-infrared tomography of buried cracks, and no claim is made that any POMEAS lens has been integrated into the referenced equipment. If the inspection requirement includes subsurface cracks, the specification has to carry a separate instrument with its own detection limit stated against the layer thickness, and the two acceptance standards have to be written separately so that a surface station is not signed off against a crack requirement it was never able to meet.

Frequently asked questions

Can a normal telecentric lens inspect cracks under ABF?

Not by assumption. A visible-light telecentric system images one object plane, so a crack buried under a build-up layer appears as a low-contrast overlay on the surface pattern, and no lens or camera choice recovers the axial discrimination. Detecting a buried crack needs illumination delivered through the volume and collection restricted to a defined scattering geometry, together with a stated detection limit against layer thickness. Validate on cross-sectioned samples before accepting any claim, including a supplier's.

Is a 3 µm minimum detectable crack the same as 3 µm dimensional accuracy?

No. A minimum detectable crack is a detection threshold under defined settings on a defined stack, and it describes whether a crack of that size can be found at all. Dimensional accuracy describes how close a measured value is to the true value for a feature that is already being measured. The two numbers come from different tests and should never be interchanged in a specification.

Can one station measure TGV diameter and RDL line width?

Not at the same magnification. A whole-panel via array needs a field of tens of millimetres, which puts the pixel scale in the tens of micrometres; a 3 µm RDL line needs a pixel scale below a micrometre. The published POMEAS range spans both ends, from 144 x 108 mm at 0.061X to 8.8 x 6.6 mm at 1.0X, but a single fixed lens cannot serve both, and a station that claims to usually delivers one of the two measurements badly.

What panel area fits a 0.061X telecentric lens?

With a 2/3 inch sensor the derived field is 144.3 x 108.2 mm, which covers a quarter of a 510 x 515 mm panel but not the whole panel in one exposure. A full-panel requirement therefore becomes a stitching or sampling plan with a defined set of measured regions, and the inspection report has to state which regions were measured. The 0.061X lens carries a 398 mm working distance, which makes carrier clearance feasible at that field size.

Next step

Send the panel drawing, the layer stack with thicknesses, the via diameters at both ends, the surface dimension list with tolerances and the smallest defect that has to be found. The first answer is arithmetic: which measurements fit which field, and which of them belong to a different instrument. For the surface station, the telecentric lens guide for semiconductor inspection covers the magnification and depth-of-field trade-offs, the note on ghost edges on glass covers the second-surface artefact that affects every transparent measurement, and the analysis of micro-hole edge blur in backlight explains what limits a small through-feature. For panel-level build-up work, the panel-level packaging metrology comparison covers line scan, white-light interferometry and film thickness as separate methods. The PMS-LTCM0061-398 product page and the telecentric lens range carry the current model data, and the contact page is where the drawing goes.