Machine vision / Semiconductor
Machine Vision in Semiconductor Manufacturing: Why It Stopped Being Optional
In semiconductor manufacturing the defects that decide your yield are smaller than a person can see, and the tolerances are tighter than a person can hold. At advanced nodes a particle under 0.1 microns can kill a die, and a misalignment of a few micrometres during lithography or dicing can waste an entire wafer. Every chip passes through hundreds of controlled steps, and a defect missed at step 40 propagates through every layer after it. Machine vision is the only gate that measures at that scale, on every unit, at line speed, and leaves an image and a number behind as evidence. This article walks the stations, the defect classes, and the yield economics that make it non-negotiable.
The numbers at a glance
01 / The physics
The defect is smaller than the inspector
Semiconductor manufacturing is the one industry where the inspection problem is settled by physics before anyone argues about budget. Producing a modern chip requires hundreds of tightly controlled steps from wafer fabrication to packaging, each stage introducing new opportunities for error, and during lithography or dicing a misalignment of just a few micrometres can render entire wafers unusable. There is no version of that sentence in which a person with a microscope is the control.
The scale gets worse the further up the node you go. At advanced nodes, particles smaller than 0.1 microns can render an entire die unusable, and a single undetected pattern deviation in photolithography propagates through every subsequent layer as a yield-destroying defect. Current AI wafer inspection platforms are built to resolve sub-micron surface defects below 1 micron, which is roughly one hundredth the width of a human hair and about two orders of magnitude below what an inspector can reliably judge at production pace.
The second problem is compounding. In electronics assembly a missed defect costs you one unit. In semiconductor manufacturing a missed defect at an early layer costs you everything spent on that wafer afterwards, and every process step after the defect adds value to material that was already scrap. That is why fabs run inspection between layers rather than only at the end: the point of detection determines how much good money follows bad.
Both facts point the same way. The gate has to be optical, automated, in line, and running on every unit rather than a sample. Everything else in this article is a consequence of those two paragraphs.
02 / The station map
Where vision sits, step by step
Machine vision is not one machine in a semiconductor plant. It is a series of gates, each with its own defect taxonomy, resolution requirement and cycle time. This is the map from wafer to tape-and-reel.
| Station | What vision does | Defect classes | Why not a person |
|---|---|---|---|
| Wafer surface, per layer | Full-wafer anomaly detection after deposition, lithography, etch and implant | Particles, scratches, spin defects, exposure problems, hot spots, contamination | Sub-micron scale, and the defect must be found before the next expensive layer is added |
| Wafer edge and bevel | Edge and sidewall inspection, including four-side direct imaging | Chipping, cracks, edge flaws, film peeling | Edge defects propagate into handling breakage later in the flow |
| Internal structure | Short-wave infrared imaging through silicon and package material | Internal cracks, voids, delamination in wafer-level packages | The defect is not on a surface. It is invisible to any visible-light method, human or machine |
| Alignment and overlay | Fiducial location and positional verification for lithography and dicing | Misalignment, overlay error, saw-street deviation | Tolerances of a few micrometres, judged in real time as a machine input rather than a report |
| Die sort | High-speed sorting with six-sided automated optical inspection before tape-and-reel | Chip-outs, cracks, contamination, ink-dot and orientation errors, sidewall damage | Volume and six faces per die. This is the station that replaces conventional manual inspection outright |
| Die attach and wire bond | Bond placement, loop profile and pad verification | Missing or misplaced bonds, tilt, epoxy bleed, wire sweep, pad damage | Feature sizes below reliable human resolution, at bonds-per-second rates |
| Bump and pillar metrology | Volumetric height measurement across the array | Bump height variation, coplanarity, missing or deformed bumps, warpage | These defects are defined by height. A flat image can infer them from shadow but cannot measure them |
| Substrate and strip | 3D AOI on lead frames, substrates and moulded strips | Mould flash, incomplete fill, lead deformation, plating defects, foreign material | Hundreds of units per strip, inspected against dimensional tolerances rather than appearance |
| Marking and traceability | Laser mark verification, OCR and 2D code reading per unit | Wrong, illegible, offset or missing marking, mixed lots | A mixed lot is a containment event, and per-unit traceability is a customer requirement, not a nicety |
| Final visual | Appearance and dimensional gate before pack-out | Cosmetic damage, handling marks, lead bend, package chips | This is the last gate before your customer's incoming inspection becomes the auditor |
Read that table as a sequence rather than a menu. Each gate protects the value added by every step before it, and the cost of an escape rises monotonically from the first row to the last. A plant that inspects only at final visual is not running ten gates cheaply. It is running one gate expensively.
03 / The yield economics
Inspection is yield management
In most industries inspection is a quality function. In semiconductor manufacturing it is the primary lever of yield, and yield is the primary driver of plant profitability. The industry-level figure is stark: undetected defects are estimated to cost the global semiconductor industry over 50 billion US dollars a year in yield loss, and traditional rule-based inspection designed for explicitly programmed defect types cannot keep pace with the expanding defect taxonomy of advanced nodes.
The plant-level arithmetic is more useful. Because die value at advanced nodes is high and volumes are large, a small percentage change in yield moves a large amount of money. Benchmark analysis of AI-based defect detection reports that a 1 percent yield gain at advanced-node die values can pay back an entire inspection platform investment within a single production month, and that moving classification accuracy from around 70 percent on rule-based systems to about 95 percent with AI vision changes the economics of yield management rather than merely improving it.
There is a second saving that rarely reaches the business case: excursion containment. Running statistical process control on defect density per wafer, rather than waiting for end-of-lot electrical test, catches an excursion at the first affected wafer before the problem propagates into work-in-process worth millions. Inspection data is not only a gate, it is an early-warning signal for the process that created the defect.
Which is the argument for treating inspection as capital equipment with a yield return rather than overhead with a quality justification. The plants that make this investment easily are the ones that measure cost of poor quality properly, as we set out in the real cost of a defect escape.
04 / The human limit
A parts-per-billion target cannot be sampled
Customer expectations in this sector have moved somewhere manual inspection cannot follow. In automotive semiconductor, 10 defective parts per million used to be a solid benchmark and the bar is now 10 defective parts per billion, driven by the roughly 1,000 integrated circuits in a modern vehicle that must all perform across a fifteen-year service life. That is a three-order-of-magnitude tightening inside one product generation.
Set that against what a person can deliver. Research from Sandia National Labs found a single inspector detects about 80 percent of defects, and two inspectors in tandem top out near 96 percent. Production studies put the working range at 70 to 80 percent, with accuracy falling 15 to 25 percent after about two hours of continuous visual work and inspector-to-inspector agreement on severity running only 55 to 70 percent. Those figures are not a criticism of inspectors. They are the specification of human vision under repetition, and we cover them in detail in your inspectors catch four in five defects.
The deeper problem is statistical rather than perceptual. A target expressed in parts per billion cannot be demonstrated by sampling, because the sample sizes required to observe single-digit defects per billion are larger than most production lots. The only way to make a claim at that resolution is to inspect every unit and keep the record. Sampling tells you about the units you looked at. Your customer is asking about the units you did not.
This is also why microscope-based final visual inspection persists longer than it should in assembly and test operations. It feels rigorous, it produces a signature, and it cannot produce a number. The moment a customer asks what else shipped from the affected lot, an image and a measurement per unit is the difference between a containment plan and a guess.
05 / Not the same as PCB AOI
Four things that change at semiconductor scale
Plenty of vision systems described as AOI are excellent on printed circuit assembly and unsuitable for die and package work. The differences are specific, and worth knowing before a vendor tells you their platform covers both.
| Dimension | PCB assembly AOI | Semiconductor vision |
|---|---|---|
| Resolution | Tens of microns is usually sufficient for component and joint features | Sub-micron, which drives optics, lighting, vibration control and data volume together |
| Defect depth | Surface and shape. X-ray is added for joints hidden under packages | Internal defects matter. SWIR imaging detects cracks, voids and delamination inside wafer-level packages |
| Geometry | 3D height on solder joints, judged against IPC-A-610 criteria | Bump, pillar and pad metrology across an array, with coplanarity and warpage as measured outputs |
| Defect taxonomy | A stable, well-documented list of component and solder defects | Expanding with every packaging generation, which is what breaks fixed-rule systems |
| Traceability | Per board, exported to MES | Per die or per unit, feeding SPC and excursion detection, not just pass or fail |
| Throughput unit | Boards per hour | Dies or units per second, with six-sided coverage on sorting stations |
| Changeover | Programme per product family, updated per revision | Recipe per device and package type, across a mix that changes with every customer ramp |
The row that catches most buyers is the fourth one. A rule-based system encodes today's defect list, and advanced packaging keeps producing defect classes that were not on it. As one industry summary of the challenge puts it plainly, advanced 3D structures often conceal internal defects that are difficult to detect, while rising investment in advanced manufacturing places greater emphasis on first-pass yield. Fixed rules and a growing taxonomy is a losing combination, and it is the reason AI classification has moved from differentiator to baseline in this sector.
06 / What to demand
Four capabilities, and one question each
Add a fifth question about false calls, because it is the operational constraint on everything above. Reported false call ratios on inspection programmes run 30 to 80 percent of total calls depending on programme quality, and every one of those calls consumes a person's time and erodes trust in the machine. The per-shift cost model is in what a 1 percent versus 5 percent false call rate costs you. Demand the figure measured over a full shift after tuning, in writing.
07 / What we do
Semiconductor vision is our specialisation
Semiconductor is the sector CODETRACE machine vision and AOI is built around, and it shapes how we work. Inspection runs on every unit rather than a sampled batch, so the height-defined defect that a flat camera keeps missing becomes a number your line can gate on. Models are trained on your own dies, packages and defect classes, not a generic library, because the defect taxonomy that matters is the one your process produces.
Deployment is on site. Systems are commissioned and tuned across Selangor and the Klang Valley, because in this sector the last stretch of accuracy comes from tuning against the real device under the real lighting on the real handler, and it cannot be done from a datasheet.
CODETRACE is a member of the NVIDIA Inception programme, and the engineers who scope a system are the ones who commission it. Bring us the die you keep failing at final visual, or the escape your customer found, and the first output is an honest assessment of whether vision can see it at all.
08 / Where to start
Start at the station that leaks
Do not start with the process flow, start with the leak. Pull the last twelve months of customer returns and internal scrap, and sort them by the station where the defect was created rather than the station where it was found. That single sort usually identifies one or two gates that would have caught the majority of your escapes, and it is a far better guide to sequencing than a vendor's coverage map.
Then characterise the defect before shopping for a machine. Is it a surface anomaly, a dimensional deviation, a height variation, or an internal structure failure? Each answer points at different optics, and the wrong answer buys a capable machine that cannot see your problem. Once the physics is settled, run the candidate system in parallel on real production for a fortnight and compare its calls against your own findings. Two weeks of parallel data replaces every claim in this article with a figure from your own floor.
For the method comparison see AOI vs AVI vs manual QC, for the capital arithmetic what an AOI machine costs in Malaysia, and for where the sector is heading and what it will demand of inspection, Malaysia's semiconductor trends. Inspection sits inside a wider factory automation plan, but in this sector it is the station that decides whether the rest of the plant is profitable.
At this scale, inspection is not quality control. It is yield management.
FAQ / Machine vision in semiconductor
Questions, answered.
01Why is machine vision essential in semiconductor manufacturing?
Because the defects that matter are smaller than a human being can see and the tolerances are tighter than a human being can hold. At advanced nodes, particles under 0.1 microns can render an entire die unusable, and a misalignment of a few micrometres during lithography or dicing can waste a whole wafer. A chip passes through hundreds of tightly controlled process steps, each one an opportunity for a defect that propagates into every subsequent layer. Machine vision is the only gate that can measure at that scale, on every unit, at production speed, and leave an image and a measurement behind as evidence.
02Where is machine vision used across a semiconductor line?
At almost every station. In the front end it inspects wafer surfaces and edges for particles, scratches, spin defects and exposure problems, and verifies pattern fidelity after etch. In assembly and test it handles die sorting with six-sided inspection, verifies die attach and wire bonding, measures bump and pillar height on advanced packages, checks lead frames, substrates and strips, reads and verifies package marking, and performs final visual inspection before tape-and-reel. Each station has its own defect taxonomy, its own resolution requirement and its own cycle time, which is why a single generic inspection recipe does not survive contact with a real line.
03Can human inspectors meet semiconductor quality targets?
No, and the arithmetic is not close. A single human inspector detects roughly 70 to 80 percent of defects under production conditions, and two inspectors working in tandem top out near 96 percent. Automotive semiconductor customers have moved the target from 10 defective parts per million toward 10 defective parts per billion, because a modern vehicle carries close to 1,000 integrated circuits that must survive fifteen years of service. A target measured in parts per billion cannot be verified by sampling or by eye. It requires 100 percent automated inspection with a per-unit record.
04How does semiconductor vision differ from PCB AOI?
Four things change. Resolution: sub-micron rather than tens of microns, which drives optics, lighting and data volume. Depth: many critical defects are internal, such as cracks, voids and delamination inside wafer-level packages, which is why short-wave infrared imaging is used alongside visible-light optics. Geometry: bump, pillar and pad height must be measured volumetrically rather than inferred from a flat image. And traceability: the record is expected per die or per unit, feeding statistical process control rather than a pass or fail log. A PCB AOI programme is a good foundation and a poor substitute.
05Who builds semiconductor vision inspection in Malaysia?
CODETRACE specialises in machine vision and AOI for semiconductor manufacturers, with deployments across Selangor and the Klang Valley. Inspection runs on every unit rather than a sampled check, with models trained on the customer's own dies, packages and defect classes rather than a generic library. CODETRACE is a member of the NVIDIA Inception programme, and systems are commissioned and tuned on site by the engineers who scoped them.