Capability / Inspection and verification
Every inspection method has a blind side, and a programme's real escape risk is set by the defects that fall into it. Optical inspection reads surfaces; transmission X-ray reads structure; neither reads behaviour. This page sets out the eight defect classes we screen for, the method that catches each, and the acceptance limit that decides pass or fail.
Defect atlas
These are the eight defect classes that account for almost every solder escape we are asked to investigate. Each card names the primary method, the limit that decides the call, and what remains exposed after that method has run.
DEF-01 / Tombstoning
A passive that stands on one termination changes height and silhouette, so optical inspection sees it reliably on 0402 and above. Escape risk rises sharply when the part sits under a shield can or inside a connector recess, where only functional test will notice the open circuit.
DEF-02 / Bridging
Exposed leads and pads show a bridge clearly, and 0.4 mm leaded pitch is well inside optical resolution. A bridge underneath a package body is invisible to optics, so the electrical test step carries the coverage there and a net-level short is what proves it.
DEF-03 / Insufficient solder
A fillet that is thin but continuous passes an optical volume check, which makes this the quietest escape in the atlas. On a heavy thermal pad the thin joint survives the bench, then cracks after a few hundred thermal cycles, so the defence is paste volume control at printing rather than inspection after reflow.
DEF-04 / Missing or misplaced
Placement faults are the cheapest defects to catch, because they exist before the joint is formed and a reflow pass has not yet been spent on them. A part shifted onto one pad but still soldered is the harder case, since it can pass both optical and electrical screening and fail a vibration or drop test later.
DEF-05 / Wrong polarity
Optical inspection resolves polarity well down to 0402 because it reads the marking rather than the solder. The escape case is a diode or electrolytic capacitor in a package where the marking is a small chamfer, and the failure then appears as a reversed rail rather than as a joint defect.
DEF-06 / BGA void
Voiding is a reflow profile and paste chemistry outcome, and it is invisible from above the package. The commercial problem is the ball that sits just inside the acceptance band: it passes by rule, then grows under thermal cycling until the joint separates within the warranty period.
DEF-07 / Head-in-pillow
The ball rests on the paste without merging, so the joint conducts under probe pressure and separates once it sees a few temperature cycles. 2D X-ray shows a joint that looks present, which is why this class is checked in three dimensions on Class 3 and safety-related programmes.
DEF-08 / Barrel fill
Fill level inside a plated hole is hard to measure optically, and a partial fill still conducts, so it escapes every continuity test and every visual check. The confirming evidence is a microsection taken from the lot, with a minimum of two locations recorded per lot at Class 3.
Method and limit
Fourteen inspection methods are available in this facility. Where a row carries a standard, that standard is the limit quoted on the inspection plan rather than a general reference.
| Method | Coverage applied | Detection limit | Governing limit or standard |
|---|---|---|---|
| Solder paste inspection | 100% of prints on production builds | Deposited volume outside 80-120% of the aperture | Aperture area ratio above 0.66 |
| Pre-reflow optical | 100% of SMT builds after placement | 0402 and above reliably; 01005 with the high-resolution lens | IPC-A-610H Class 2 target condition |
| Post-reflow optical | 100% of all SMT builds | Fillet length and presence; judgement drops below 0.35 mm pitch | IPC-A-610H Class 2 and Class 3 |
| Two-dimensional X-Ray | Every BGA and QFN, plus a sample on legacy through-hole work | Ball presence and bridging; void depth is not measurable | IPC-7095 reference for voids |
| Three-dimensional CT X-Ray | Every BGA on Class 3 and safety-related programmes | Void percentage per ball and head-in-pillow separation | Under 25% Class 2, under 10% Class 3 |
| Bench visual inspection | 100% at Class 3, sampling at Class 2 | Workmanship above roughly 0.4 mm leaded pitch | IPC-A-610H, CIS-certified inspectors |
| UV coating inspection | Every coated assembly | Coverage and keep-out compliance, not adhesion or cure state | IPC-CC-830B qualification reference |
| Microsection | Minimum 2 locations per lot | Barrel fill, annular ring and intermetallic thickness | IPC-6012 Class 3: 25 um minimum plated copper |
| Ionic contamination | Once per lot by ROSE or ion chromatography | Residue expressed as NaCl equivalent per square centimetre | 1.56 ug Class 2, 0.78 ug Class 3 |
| Dye-and-pry | Sample after thermal cycling on BGA programmes | Cracked balls and non-coalesced head-in-pillow joints | IPC-9708 test method |
| In-circuit test | 100% where the volume band justifies a fixture | Continuity, shorts and value against the netlist | CAD netlist comparison |
| Flying probe | 100% on high-mix and prototype volumes | The same electrical facts, without fixture pressure on the joint | CAD netlist comparison |
| Functional test | 100% where the test plan specifies it | Behaviour at one defined operating point | Customer limit table with stated conditions |
| Hipot on coated assemblies | 100% on mains-connected medical and energy builds | Dielectric breakdown or flashover across the barrier | 4000 V AC for 1 minute at 2xMOPP, 250 V |
How inspection is run
An optical recipe programmed from a known-good assembly inherits that assembly's defects. We build the program from the CAD placement data so the expectation is generated rather than measured, and then confirm it against the first article. Library parts carry their own inspect windows, which keeps a 01005 passive and a 33 mm connector from sharing one tolerance. Every recipe revision is versioned, so the images retained against a serial number can be traced to the program that judged them.
The tempting fix for a noisy optical line is to widen the window until the false calls stop, which quietly widens the escape band as well. We review the false-call Pareto weekly and change the library part or the lighting rather than the accept threshold, and any threshold change is recorded as a recipe revision with a reason. On a Class 3 programme the accept threshold is not a tunable at all, because the rejection band widens from 2-5% at Class 2 to 8-15% at Class 3 and the physics does not move with the number.
When a defect reaches the field, or is caught late by functional test, the image set for that serial number is reopened and the recipe is asked why it passed. The answer is usually a lighting angle, a library window or a keep-out that hid the joint, and it becomes a permanent change rather than a note. That loop is what holds the published field return rate at 0.06% while the mix of packages on the line keeps getting denser and smaller.
Four-way coverage
The four methods answer different questions about the same joint. Presented together, the table shows why they are additive rather than alternative, and why removing one moves cost into returns rather than into savings.
| Comparison | Optical AOI | X-Ray | Visual inspection | Functional test |
|---|---|---|---|---|
| Sees | Surface features within line of sight | Internal structure by transmission | Whatever the eye resolves at the bench | Electrical behaviour against limits |
| Blind to | Under a BGA, can, body or thermal pad | Electrical function and marginal wetting | Features below 0.4 mm leaded pitch | Anything structural that still functions |
| Coverage here | 100% of SMT builds | Every BGA and QFN | 100% at Class 3, sampling at Class 2 | 100% where the test plan specifies it |
| Cycle time | 5-15 seconds per board | 30 seconds to several minutes per board | 30-120 seconds per board | 30 seconds to 5 minutes per board |
| Cost per board | $0.20-0.50 | $1-5 in 2D, $5-20 in 3D | $0.10-0.25 | $1-25 depending on test level |
| Catches | Bridging, tombstoning, polarity, missing part | Void, ball integrity, barrel fill, head-in-pillow | Workmanship, labelling and mechanical fit | Wrong value in an identical package, and dead rails |
| Reference cited | IPC-A-610H Class 2 and Class 3 | IPC-7095 void acceptance | IPC-A-610H with CIS-certified inspectors | The customer limit table and its conditions |
| Escape left open | Hidden joints and thin but continuous fillets | Electrical failure with intact structure | Fine geometry and sub-surface defects | Marginal joints, drift and wear-out |
Macro record
One transmission image and two optical macros, taken at the magnification the recipe actually judges at rather than at a marketing scale.
Buyer questions
Optical inspection needs a line of sight, so it is blind under a ball grid array, under a shielded can, under a connector body and beneath a large thermal pad. It also struggles with a joint that is thin but continuous, because a fillet that is present in the wrong volume still looks like a fillet from above. That is a hardware boundary rather than a camera problem: the defect classes that escape an optical pass are covered by transmission X-ray, by in-circuit test, or by functional test, and the coverage report names which one is doing the work.
Under 25 percent of ball diameter at Class 2 and under 10 percent at Class 3, measured per ball rather than per package, which is the IPC-7095 reference. The number only means something with the measurement method attached, because 2D and 3D X-ray report different void figures on the same joint. We state the method beside the limit in the inspection plan, and we retain the 3D CT images per serial number on Class 3 and safety-related programmes.
Yes, and the two usual reasons are head-in-pillow and wrong value in an identical package. Head-in-pillow is a ball that touches the paste without coalescing, so it conducts on the bench and separates after a few thermal cycles; it is why the joint is checked with 3D CT rather than 2D X-ray on Class 3 work. A wrong value in the same package is optically identical to the correct part, so only an electrical test with a written limit catches it. The published field return rate of 0.06 percent is the number that tells us how well the strategy is holding.
Seven fields. A process engineer replies within one working day.
Seven fields. A process engineer replies within one working day.
Seven fields. A process engineer replies within one working day.