HBM and WLP Inspection: Separate 3D Metrology from 2D Defect Review

Evidence disclosure: this page is an independent engineering guide informed by public company information from OPTIVE. It is not a customer case, an installed-base claim, an endorsement or a verified compatibility statement. The camera sensor, the objective part numbers and any purchasing relationship with POMEAS are UNKNOWN.

An HBM stack and a wafer-level package are both judged on two different questions at once: how tall is it, and does it look right. The first question is quantitative and needs a calibrated height reference somewhere in the chain. The second is qualitative and needs contrast, colour and enough magnification to classify a defect. Procurement collapses the two into "inspection optics" and then discovers, usually at acceptance, that the station which measures coplanarity well cannot see the defect that matters, and the station that images the defect beautifully has no traceable height output at all.

The separation is not a matter of buying a better lens. It comes from the measurement principle: a 2D imaging chain produces a lateral coordinate per pixel and nothing else, while a height chain produces an axial coordinate against a reference. This page sets out what each chain has to contain, does the arithmetic that decides magnification, and states where a visible-light review station stops and a different instrument has to take over.

A bonded stack is not a flat sample

A finished HBM stack is a composite: a logic base die, stacked DRAM dies, through-silicon vias, micro-bumps or hybrid bonds between them, underfill, and often a mould cap. The top surface that a review camera sees may be silicon, mould compound or a lid, and each of those returns light differently. The features that matter are at three scales simultaneously: die-level alignment marks measured in tens of micrometres, bump arrays with pitches in the tens of micrometres, and particle or scratch defects that can be a few micrometres across.

Wafers add a second complication. A 300 mm wafer is never flat to the depth of field of a high-magnification objective, so either the stage tracks focus per field or the field has to be small enough that the local tilt is tolerable. Both of those are mechanical decisions that constrain the optics, and neither of them is captured by a magnification number on a datasheet.

Three optical jobs, not one station

The first job is overview and localization: find the die, find the field, decide where to look. It wants a large field, a generous depth of field, and tolerance for height variation, and it does not need micrometre resolution. A 1.0X bi-telecentric lens with an 18.4 mm object field is a typical instrument here, because a telecentric field keeps the image scale constant as the stack height changes across the wafer.

The second job is high-magnification review: classify the defect, measure the feature, decide pass or fail. This is where numerical aperture decides what exists in the image at all, and where the depth of field becomes small enough that focus has to be actively maintained. A 5X or 10X plan apochromatic objective is the usual class, and the choice between them is an NA decision before it is a magnification decision.

The third job is height: bump coplanarity, die tilt, mould thickness, via reveal. No 2D optical chain produces this, however fine its pixel. It needs a measurement principle with an axial reference, which in practice means confocal, white-light interferometry, or a laser triangulation/displacement sensor with a calibrated scale. The 2D review station can be excellent and still contribute nothing to this number.

What the external OEM actually discloses

OPTIVE publicly describes 3D metrology optics for HBM, WLP and 3DIC, large-format 3D optics development, high-speed macro inspection and automated defect review. That is a description of a company's capability set, and it is useful as a map of which problems the market treats as related. It does not establish which camera sensor is used, which objective part number is inside, or that any POMEAS optic is purchased or interchangeable. Those fields remain UNKNOWN, and a specification written as if they were known would be inventing facts.

Resolution arithmetic: NA first, pixels second

Start with the optics. For a diffraction-limited system in air, the resolvable detail is on the order of 0.61 times the wavelength divided by the numerical aperture. Taking a nominal 0.5 µm wavelength for the arithmetic, the published POMEAS objective apertures map to the following limits.

Then check the sampling, because it is usually the number that gets quoted and the one that misleads. With a 1.1 µm camera pixel and 5X magnification, the object sampling is 1.1 divided by 5, or 0.22 µm per pixel (derived), so a 2 µm feature spans about nine pixels. Nine pixels sounds comfortable. Against a 2.18 µm diffraction limit, those nine pixels are sampling an image that the objective has already blurred: the pixel scale is about ten times finer than the optical limit. Increasing the camera resolution further changes nothing at all.

That oversampling is not useless — sub-pixel edge fitting and centroid measurement benefit from it — but it is not free capability. A specification that lists the pixel scale without the NA, or the camera resolution without the objective, is describing half of an imaging chain, and in HBM review work it is usually the less important half.

Depth of field sets the mechanics, not the optics

Depth of field falls roughly with the square of the numerical aperture, so the objectives that resolve a 2 µm defect are the ones with almost no tolerance for height. The published POMEAS figures show the trend across the range: 18.75 mm at 0.16X, 1.00 mm at 1.0X with an 8.39 µm resolution, 0.28 mm at 10X for the 1.2 µm class, and a depth of field on the order of a couple of micrometres at NA 0.55 by the same estimate. On a stack whose top surface varies by tens of micrometres across a die, a 5X or 10X field cannot be imaged without focus tracking.

The practical consequences are unglamorous and decisive. If the stage has no Z axis, the achievable magnification is capped by the flatness of the part. If the stage has a Z axis but the focus search is slow, the review throughput is set by the search and not by the camera. If the part sits on a carrier whose thickness varies, the variation enters the measurement as a focus error rather than as a dimension error, which is worse, because it is invisible in the data.

Height is a different instrument chain

A height result has to be traceable to a standard. That means a sensor with a calibrated axial response, a reference surface measured in the same session, and an uncertainty budget that includes the sensor, the stage, the temperature and the algorithm. A 2D review station that reports a focus position is reporting a machine coordinate, not a height, and the two are only convertible if the whole chain has been characterised.

POMEAS covers the height side with separate product families: white-light interferometry for areal surface topography such as bump and pad geometry, and confocal displacement sensing for inline height and step measurement. These are instruments with their own specifications and their own calibration requirements, not accessories to a review camera. When the requirement is coplanarity across a die, the honest architecture is a 3D sensor measuring a sampled grid plus a 2D station reviewing the defects that the grid flags, with the two data sets joined by die and field coordinates rather than by one instrument pretending to do both.

Architecting the split

A motorised zoom is worth considering for the navigation role specifically because it changes field without changing stations: a 0.68X-5.0X range with an 80 mm working distance covers overview and mid-magnification in one optical path. The condition is that the zoom position is repeatable and calibrated, because a zoom lens whose magnification is only known to a few percent is a viewing aid, not a measuring instrument. For applications that need telecentric behaviour through the zoom, the 0.5X-1.0X telecentric macro zoom keeps distortion below 0.02 percent and the working distance in the 75-106 mm band, which is the class to look at when scale stability matters more than range.

Engineering acceptance matrix

Before any of this is quoted, six inputs have to exist on paper: the camera active area and pixel pitch, the objective NA and working distance, the illumination wavelengths and geometry, the feature size and defect height of interest, the wafer or panel flatness and warpage, and the required uncertainty with the calibration artefact that will demonstrate it. Two further items are usually forgotten: the review speed per field, which sets the stage and focus requirements, and the ownership of the coordinate frame that joins 2D review results to 3D height results.

Validation should cover the full field and the corners, not only the centre, because field curvature and distortion both grow outward and a centre-only acceptance test passes optics that fail in production. Lateral and axial metrics should be recorded separately and never averaged into one "accuracy" number, since they come from different physical references and fail for different reasons.

Frequently asked questions

Can a zoom lens replace a 3D metrology objective?

No. A zoom lens changes magnification and field; it does not create an axial measurement reference. Replacing a 3D objective means replacing the height measurement principle, with its own calibration, spot size and uncertainty budget. A zoom can serve navigation and 2D review in the same station, which is a useful simplification, but the height number still has to come from a confocal, interferometric or triangulation sensor calibrated against a standard.

Does a smaller object pixel mean better height accuracy?

No. Object pixel size is a lateral sampling figure. Height accuracy depends on the axial measurement principle, its calibration and its stability, and on the flatness reference used during the measurement. A camera with a 0.11 µm object pixel at 10X has no axial output at all unless the focus axis is calibrated, and even then the result is a machine coordinate rather than a traceable height.

Which magnification is needed to review a 2 µm defect?

Work from the aperture rather than the magnification. At a nominal 0.5 µm wavelength, the published POMEAS apertures give a diffraction limit of about 2.18 µm at NA 0.14 and about 1.09 µm at NA 0.28. A 2 µm defect is therefore at the limit at 5X and roughly two resolution elements across at 10X, which is where it becomes classifiable rather than merely visible. Confirm on real samples with the production illumination before committing.

Can one lens cover overview and defect review?

Only with a zoom, and only with accepted trade-offs. A 0.68X-5.0X motorised zoom covers both fields in one optical path at an 80 mm working distance, but the scale at intermediate positions must be calibrated if measurements are taken there. A fixed 1.0X bi-telecentric lens is the better overview instrument because its scale does not move with height, and a fixed 5X or 10X objective is the better review instrument because its aperture makes the defect resolvable.

Next step

Send the stack cross-section, the defect list with sizes, the height and coplanarity requirement, the camera sensor and the flatness of the incoming wafer or panel. The first reply separates the requirements into the three chains above, so that the quotation prices the height sensor and the review optics as the different instruments they are. For the review side, the telecentric lens guide for semiconductor inspection covers scale stability and depth of field, and the SiC wafer comparison of zoom against objective with laser autofocus covers the focus-tracking decision that dominates high-magnification review. For panel-level build-up work, the panel-level packaging metrology comparison sets line scan, white-light interferometry and film thickness side by side, and the wafer-level flash measurement guide covers the case where the sample is flat enough for a single calibrated exposure. The PMS-MPO5 5X objective, the PMS-MPO10 10X objective and the PMS-10VP110C111T 1.0X bi-telecentric lens carry the current model data; the contact page is where the drawing goes.