A Japanese precision-parts manufacturer chose 2D AI vision inspection over the 3D structured-light system that their evaluation committee had initially specified — at approximately one-third the capital cost, with 2 days of on-site setup to production readiness. The decision was not primarily about cost. It was about the defect class the line was actually failing on: surface micro-cracks and edge chipping on machined precision components, which the proposed 3D system resolved as surface noise rather than defect signal. The same AI model that caught the micro-cracks the 3D system missed was running in production within the same week.
Micro-cracks and sub-surface defects in semiconductor packages and precision machined components produce the most expensive field failures — not because they are hard to detect in theory, but because standard single-shot AOI and visible-light inspection miss them systematically. The field failure arrives months after shipping, by which point the defect population in the field cannot be bounded without costly recall action.
What micro-cracks are and why they escape standard inspection
A micro-crack is a fracture that does not propagate through the full cross-section of the component at the time of inspection. In semiconductor packaging, the most common micro-crack morphologies are:
Die corner cracking — at the edge of the silicon die, where dicing stress or thermal cycling during burn-in creates crack initiation sites. These cracks are typically 5 to 50 micrometres in length and may have sub-micrometre aperture at the surface. A single-shot overhead AOI camera with standard co-axial illumination sees the die corner as a bright-edged feature; a crack at the corner is within the existing edge contrast and does not produce a distinct additional signal.
Package delamination — at the interface between the mould compound and the die or leadframe in plastic-encapsulated packages. This is a sub-surface defect: the delamination initiates at the interface, invisible from the external surface. Initial delamination produces no visible surface feature. Propagation during temperature cycling in field operation eventually leads to bond wire fatigue failure or die cracking at the delamination front.
Substrate micro-cracking in precision machined components — on bearing races, valve seats, precision fasteners, and hydraulic components, surface micro-cracks from grinding, EDM, or heat-treatment stresses concentrate at stress-riser geometries (corners, cross-holes, thread roots). These cracks are characteristically sub-surface in origin — they initiate below the surface finish from residual stress — and have surface expressions of less than 10 micrometres in width.
The common characteristic across these morphologies: they are present at the time of manufacture, they do not produce field failures immediately, and they propagate to failure under the thermal cycling and mechanical stress of service conditions — typically over a period of 6 to 24 months. By the time field failures cluster, the original production batch is distributed across multiple customers and the defect population cannot be bounded without full field recall.
Quality engineers who work precision-component lines describe the discovery process bluntly: the first indication that a batch had micro-crack escapes is a warranty cluster. The components passed AOI, passed functional test, shipped, and worked for a year before the thermal cycling of their application propagated the crack to failure. The batch is long gone. The cost is not the component; it is the field investigation, the containment, and the customer conversation about why a "passed inspection" part failed in service.
Why standard AOI misses these defects
Single-shot AOI is designed around the defect classes that produce detectable contrast under its illumination configuration. Standard overhead co-axial illumination — the most common AOI illumination geometry — produces good contrast for solder joint presence, component polarity markings, and pad coverage. It is a poor geometry for micro-crack detection for three reasons.
Illumination angle. A micro-crack with a sub-micrometre surface aperture produces its detectable signal only when illuminated from an angle where light enters the crack and is scattered or absorbed — typically low-angle grazing illumination at the crack's orientation axis. Co-axial or diffuse illumination reflects off the surrounding surface; the crack signal is buried in the background. The same crack that is invisible under co-axial light is clearly visible under 10° to 15° grazing illumination at the correct rotational angle.
Single image capture. A crack oriented perpendicular to the illumination angle produces no contrast signal, regardless of illumination intensity. Single-shot inspection captures one image. Multi-angle inspection captures images from multiple illumination directions, rotating through the grazing-angle orientations that different crack orientations require. The defect signal that is invisible in any single image becomes unambiguous when images from complementary angles are compared.
Resolution and field of view trade-off. On a die with 5 mm of inspectable perimeter, a micro-crack at 10 micrometres in width requires a camera resolution better than 5 micrometres per pixel to resolve reliably. Standard AOI cameras set their resolution for the full board FOV, which on a multi-component PCB is 20 to 50 micrometres per pixel. A precision component with smaller inspection area can be imaged at higher resolution, making crack detection tractable with the appropriate camera specification.
Thermal imaging as a complement for sub-surface delamination
For package delamination and sub-surface crack networks in sealed components, optical imaging cannot access the defect directly. The complement modality is lock-in thermography: the component is subjected to a brief thermal stimulus (resistive heating, flash lamp, or ultrasonic excitation), and the thermal response is captured by an infrared camera. Delaminations and crack networks conduct heat differently from intact material, producing thermal contrast above the defect location.
Lock-in thermography does not replace optical surface inspection — it addresses the sub-surface defect class that optical inspection cannot see. The two are complementary, not competitive. A production inspection station for high-reliability semiconductor packages may include both: optical multi-angle inspection for surface micro-cracks and edge defects, and thermographic inspection for delamination screening.
The practical consideration: thermographic inspection cycles are longer than optical inspection cycles (a typical lock-in sequence is 3 to 10 seconds per component versus sub-second for optical), and the station cost is higher. For high-unit-value components where the field-failure cost is in the range of $500 to $5,000 per escaped unit, the cycle time and capital cost are justified by the escape cost curve. For high-volume, low-unit-value components, thermographic screening is typically applied to statistical samples or to lots flagged by optical inspection, not to every unit.
The escape cost argument: why 99% detection is not the number to optimise
Vendors quote detection rates. Buyers should be asking about escape cost curves.
A detection rate of 99% sounds high. On a precision component production run of 50,000 units with a defect rate of 500 ppm (0.05%), the defect population is 25 units. At 99% detection, the expected number of escaped defective units is 0.25 per run — less than 1 unit escaping per run, on average. This sounds acceptable.
The escape cost calculation changes the picture.
| Defect rate (ppm) | Units in batch | Defects in batch | Detection rate | Expected escapes | Field failure cost per unit | Expected escape cost |
|---|---|---|---|---|---|---|
| 500 | 50,000 | 25 | 99% | 0.25 | $200 | $50 |
| 500 | 50,000 | 25 | 99% | 0.25 | $2,000 | $500 |
| 500 | 50,000 | 25 | 99.9% | 0.025 | $2,000 | $50 |
| 2,000 | 50,000 | 100 | 99% | 1.0 | $2,000 | $2,000 |
| 2,000 | 50,000 | 100 | 99.9% | 0.1 | $2,000 | $200 |
Two things stand out. The cost of a 1% detection gap is entirely determined by the field-failure cost per escaped unit — a number that varies by application from tens to thousands of dollars. And the marginal value of improving detection from 99% to 99.9% is only justified by the specific escape cost in your application. A component going into a consumer electronics product with a $5 warranty claim has different inspection economics from a precision fastener going into an aerospace assembly with a $3,000 non-conformance cost.
The claim to chase is not the detection percentage alone. The question is: what does one escaped defective unit cost you, and what is your current defect rate? Those two inputs determine how much detection accuracy is worth buying — and whether the detection specification your vendor is quoting addresses the right number.
Multi-angle imaging in production: what it looks like at the station level
A multi-angle inspection station for semiconductor and precision components consists of:
- A primary overhead camera for component presence and gross defect detection
- Two to four grazing-angle illumination modules at different rotational positions (typically 0°, 90°, 135°, and 180°, or a rotating illumination dome)
- A separate camera for each grazing angle, or a single camera with synchronised sequential illumination
The capture sequence: the component arrives at the station, overhead image is captured, then four grazing-angle images are captured in rapid sequence. Total capture time for a five-image sequence is 80 to 150 milliseconds, depending on illumination switching speed and camera frame rate. This is compatible with production throughput at 400 to 600 components per minute on dedicated precision-component lines.
The inspection model receives all five images as input and classifies the component against trained defect classes: die corner crack, edge chip, surface micro-crack, delamination edge signature. The model's per-image decision is aggregated across angles: a defect that appears in two or more angle images with consistent morphology is a confirmed flag; a feature that appears in one image only with no corroboration in complementary angles is reviewed separately. This multi-image voting logic reduces false calls on surface reflections and tooling marks that produce localised contrast under a single illumination angle.
What the precision-parts deployment involved: evaluation to PPAP
The 3D system the evaluation committee had originally specified resolved the surface micro-cracks as texture noise — within its measurement uncertainty, the crack signatures fell below the detection threshold. HyperQ AI Vision, using multi-angle grazing-illumination capture and a model trained on confirmed micro-crack morphologies from the client's own production samples, detected the same defects at a rate the quality team accepted in their PPAP submission. That model was not trained on generic crack reference data. It was trained on the specific morphology the line was producing: cracks at specific orientations, apertures, and substrate finishes from that component family.
Setup time from arriving on site to a production-ready model was 2 days. The 3D system's installation and validation cycle had been quoted at 6 to 8 weeks.
The capital cost difference — approximately one-third of the 3D system's price — did not drive the decision. The detection performance on the actual defect class did. But in semiconductor and precision-component manufacturing, where inspection station count scales with the number of distinct component families on the line, the capital cost difference compounds. A 3D system at full cost per station is a different capital conversation from a 2D AI station at one-third that cost, particularly when the 3D system's dimensional measurement capability is not the inspection requirement.
For context on how AI escape analysis compares with AOI as a complementary layer in electronics, see AOI vs AI escape analysis in electronics: when to add a second-pass vision model, which covers the second-pass architecture in detail.
When 3D inspection and other modalities still win
Multi-angle 2D AI inspection is not always the correct answer for sub-surface defect detection.
True sub-surface delamination screening. Where the defect is entirely sub-surface — package delamination with no surface expression, internal voids in cast or moulded components — optical inspection cannot access the defect regardless of illumination configuration. Thermographic inspection, scanning acoustic microscopy (C-SAM), or X-ray are the appropriate modalities.
Dimensional metrology requirements. Where the inspection requirement includes absolute dimensional measurement — diameter within ±1 micrometre, flatness within ±0.5 micrometre — 3D structured-light or white-light interferometry provides traceable dimensional data that 2D AI inspection cannot. The HyperQ-style AI approach is not a metrology system. It is a defect classification system.
Very high defect frequency at low line speed. Where a line's defect rate is high enough that every component requires detailed multi-angle inspection and throughput is below 100 components per minute, scanning acoustic microscopy may provide better information density per component at acceptable cycle time.
Frequently asked questions
What is the minimum crack width that 2D multi-angle AI inspection can reliably detect? Under optimal grazing-angle illumination, cracks with surface aperture at 5 to 8 micrometres are detectable when the camera resolution is 2 to 3 micrometres per pixel. At the apertures typical of post-dicing die corner cracks (10 to 30 micrometres) and grinding-induced substrate cracks (8 to 20 micrometres), detection with correctly configured grazing illumination is reliable. Cracks below 3 micrometres in surface width require scanning electron microscopy or acoustic microscopy to detect consistently — optical inspection at any illumination configuration reaches a physical resolution limit at that scale.
Does the AI model need to be retrained when a new component family is introduced? Not always from scratch. The pre-trained model library includes crack morphology classes that generalise across metal component families and semiconductor package types. Where a new component family has similar surface finish and defect morphology to an existing library entry, the library model may be used directly with validation. Where the surface finish, geometry, or defect morphology differs substantially, a client-specific training set from the new component family is required. The low-data training architecture reaches production readiness with 1,000 labeled images per defect class.
How does multi-angle illumination handle highly reflective surfaces? Specular surfaces (polished metal, plated finishes) scatter grazing illumination differently from matte surfaces. The illumination wavelength and polarisation can be adjusted to reduce specular glare and improve crack-edge contrast. For mirror-polished bearing races, dark-field illumination — where the background is dark and surface features scatter light toward the camera — often produces better crack contrast than bright-field configurations. This is a setup parameter determined during the trial evaluation, not a fixed system constraint.
What quality records does the inspection station produce for IATF 16949 or AS9100 audit purposes? The inspection station produces per-component records: component ID, timestamp, inspection outcome (pass/fail/uncertain), defect class if flagged, and image archive. For IATF 16949 audit purposes, the inspection record chain from component serial number to inspection outcome to disposition is available as a data export. For AS9100 and aerospace traceability requirements, the image archive provides the visual evidence that a component was inspected and the specific defect determination that supports disposition decisions.
The escape cost for an undetected micro-crack in a precision component is not the cost of the component. It is the cost of the field failure — plus investigation, containment, rework, and relationship capital. Send us a confirmed defect sample from your current production: 50 to 100 images of confirmed micro-crack morphologies from your line, plus your current inspection station configuration. Within 2 weeks we will specify a multi-angle model with detection performance data against your actual defect class, at no contract obligation until performance is confirmed on your production samples.
Send confirmed defect samples for a micro-crack model specification