Structured light scanning in the forge and foundry

2026-08-24 · expanded 2026-09-08 · long read · sources and confidence marks throughout

Forging and casting is where industrial structured light grew up, and the connecting rod explains why better than any other part. Structured light scanning is a 3D metrology method: a projector casts known light patterns onto a part, cameras record how those patterns deform across the surface, and triangulation turns that deformation into a dense point cloud. As it leaves the die, it lives in a world of half-millimetre tolerances, mismatch and die wear — full-field territory. After machining, it lives at tens of microns of bore roundness — hard-gauge territory. One part, two metrological lives, and the boundary between them is exactly where optical scanning earns its keep or doesn't. This report starts with how the instrument actually works — patterns, triangulation, calibration — and then follows the part from the die to the gauge.

How this report was built. Findings were researched from primary sources — standards documents, peer-reviewed studies, vendor technical documentation, foundry trade press — and each carries its confidence honestly: claims marked reportedly or vendor-sourced come from trade press or marketing material I could not independently verify; unmarked claims trace to the sources listed at the end. The 3D models are AI reconstructions for illustration — they are not scans of real parts, and no dimensional claim rests on them.

1 · What structured light scanning actually is

Before the foundry argument, the mechanism — because most of what follows only makes sense once you know what the instrument is doing. A structured light 3D scanner is three things pointed at the same place: a projector, one or two cameras, and a calibration that ties them together. The projector throws a known sequence of light patterns onto the object surface — parallel stripes most often, sometimes a grid. The cameras, mounted at a fixed angle to the projector, photograph the object with those patterns lying across it. Because the surface is not flat, each projected pattern arrives at the camera bent: the deformed pattern carries the object's shape written into it. Software decodes that deformation and, by triangulation, computes three dimensional coordinates for every camera pixel that saw an identifiable piece of the pattern.

Triangulation is the whole trick, and it is ordinary geometry. The projector knows which stripe it emitted. The camera sees where that stripe landed. Projector centre, camera centre and surface point form a triangle whose baseline — the fixed separation between projector and camera — is known from calibration, and whose two angles are read off the projector column and the camera pixel. One triangle per pixel, solved a few million times, and the output is a point cloud: a set of measured points on the object's shape — one for each camera pixel that decoded successfully, so it arrives on the camera's own regular lattice rather than as a scatter — which meshing then turns into a surface and, eventually, a scanned model — a digital model of the real part, ready to compare against CAD.

Two physical consequences follow immediately, and both matter on a foundry floor. First, structured light scanners capture entire surfaces at once rather than point by point — the structural difference from a touch probe, and the reason full-field deviation colormaps exist at all. A tactile CMM measures the points you told it to measure; a structured light system measures everything inside the measuring volume and lets you decide afterwards what to ask of it. That is what makes a scan a re-interrogatable record rather than a report. Second, the measurement is non-contact: nothing deflects a thin wall, nothing has to be fixtured to resist probing force, and a sand-cast surface that would wreck a stylus tip is simply photographed.

A third consequence is about what the instrument gets used for rather than how it works. Because the sensor records shape without knowing what the shape is for, the scanner that inspects a casting is the same one that does reverse engineering — which in a foundry usually means recovering geometry from a legacy pattern whose drawings are long gone. That job is a different report. This one concerns parts that leave a die or a mould with a drawing already attached.

Structured light, laser, and LiDAR are not the same instrument

These families get conflated constantly, so, plainly: structured light projects a two-dimensional pattern across an area and triangulates. Laser scanners of the triangulating kind — including the blue-laser portables discussed in section 7 — sweep a single stripe or a point across the surface and solve the same triangle, trading area coverage for tolerance of shiny and awkward surfaces. LiDAR and time-of-flight instruments do not triangulate at all; they time a light pulse, which is the right approach at room and building scale and wrong by orders of magnitude for part metrology — they resolve millimetres to centimetres where a forging tolerance is tenths of a millimetre. Photogrammetry uses no projector, deriving shape and scale from images taken from various angles, and is the standard partner for tying many scans of large objects into a single frame. So if the question is whether you need LiDAR to inspect a casting: no. Different range regime, different accuracy class, different job.

2 · Patterns, calibration, and what accuracy and resolution mean on a spec sheet

Why a scan takes multiple frames

The decoding problem is one of identification: for a given camera pixel, which projected stripe am I looking at? One stripe at a time is unambiguous but slow, since it must be swept across the whole surface. Many stripes at once is fast but ambiguous, because every stripe looks like every other stripe. Structured light patterns exist to resolve that ambiguity, and the field's solutions are a study in trading frames against certainty.

The classical answer is temporal coding with binary or Gray codes. The projector shows a sequence of black-and-white stripe patterns, each twice as fine as the last; each pixel records a run of light/dark readings, and that bit sequence is a binary address identifying which projector column illuminated it. Gray codes are preferred over plain binary because adjacent code words differ in exactly one bit, so a misread at a stripe boundary — where a pixel straddles the edge — shifts the answer by one column instead of, potentially, half the projector. Ten patterns address 1,024 columns — and ten is a floor, not a total: industrial sequences usually project each pattern together with its photographic inverse, roughly doubling the count, so that each pixel is thresholded against its own illuminated and unilluminated brightness rather than against a global guess. The price is that the part must hold still for every one of those frames.

Phase shifting then refines the answer to sub-pixel precision. Instead of hard-edged stripes, the projector shows a smooth sinusoidal fringe, then shifts it by a known fraction of a period — three or four steps is typical — and the intensities each pixel records across those frames solve for the local phase. Phase is continuous, so it interpolates between stripes and recovers fine details that a binary code alone would quantise away. Its weakness is that phase repeats every fringe, leaving an ambiguity of whole periods. The standard industrial recipe is therefore both together: Gray codes to establish which fringe you are in, phase shifting to locate the point precisely within it. That combination is why a metrology-grade scan is a burst of multiple frames rather than a photograph — and why moving parts and dynamic scenes require single-shot pattern strategies that buy one exposure at the cost of precision. Decoding all of it is a pattern recognition problem borrowed wholesale from computer vision.

The choice of light matters for the same reason a photographer chooses a flash. Blue light systems project a narrow band of the spectrum and put a matching band-pass filter on the camera, so shop ambient light is rejected and the sensor sees mostly its own projection — which is why blue-LED units work under factory lighting where older white light systems wanted a darkened room. The same trick is what lets the in-line system in section 6 measure a forging that is glowing at 900 °C, for reasons set out there. Nothing about blue is magic: it is a shorter wavelength and a narrower band, chosen for contrast against everything else in the room, and infrared light is the band you would choose if the thing you needed to see through were different.

Calibration is the accuracy

None of the triangulation means anything without the geometric relationship between camera and projector. Calibration measures the intrinsic parameters of each camera — and of the projector, treated as a camera running backwards — focal length, principal point, lens distortion, plus the extrinsic parameters that place them relative to one another, usually by capturing a certified artefact from various angles. Calibration error propagates directly into the reconstructed surface: a scanner is exactly as good as the geometry it believes it has. That is why serious systems re-calibrate on a schedule, after transport, and whenever ambient temperature moves — and why the NIST finding in section 10, that standard length tests do not reliably catch every systematic calibration error, is more unsettling than it first sounds.

A worked example: point spacing is not accuracy

The single most common misreading of a scanner datasheet is treating point spacing as accuracy. They are different quantities, related only loosely. Here is the arithmetic, on a hypothetical sensor — the numbers are invented for illustration, but every step is shown so you can run it against your own.

Setup (hypothetical). A stereo pair of 4,000 × 3,000 pixel cameras — 12.0 megapixels each — configured over a measuring volume 400 mm wide by 300 mm high.

Step 1 — point spacing. Across the width, 400 mm ÷ 4,000 px = 0.100 mm per point. Down the height, 300 mm ÷ 3,000 px = 0.100 mm. The two agree because this measuring volume matches the sensor's 4:3 aspect ratio; when a quoted volume does not, the spacing differs by axis and the coarser one governs.

Step 2 — points per scan. One point per pixel gives 4,000 × 3,000 = 12,000,000 points from a single exposure sequence. A nine-megapixel camera, on the same logic, yields about nine million.

Step 3 — change the volume. Refit the same cameras over a 100 mm × 75 mm volume: 100 mm ÷ 4,000 px = 0.025 mm. Four times finer spacing (400 ÷ 100 = 4), over one-sixteenth of the surface area — (400 × 300) ÷ (100 × 75) = 120,000 ÷ 7,500 = 16. Resolution and coverage trade against each other exactly, and no setting escapes it.

Step 4 — now the part people get wrong. Suppose per-point range noise is σ = 20 µm, a realistic figure on an as-cast skin. Fit a plane to a 20 mm × 20 mm patch at 0.100 mm spacing. Across one side, 20 mm ÷ 0.100 mm = 200 spacings, hence 201 points; the patch holds N = 201 × 201 = 40,401 points. If those N errors are statistically independent — hold that condition, it does the real work here — the standard uncertainty of the fitted plane's position falls as σ ÷ √N = 20 µm ÷ √40,401 = 20 µm ÷ 201 = 0.0995 µm, call it a tenth of a micron.

Reading. A sensor whose individual points scatter by 20 µm can, in principle, locate a fitted plane to a fraction of a micron. That is the honest reason a fitted feature gets quoted tighter than per-point noise, and it is why point spacing and accuracy are different quantities. It is not, by itself, why the certificate in section 10 quotes the MPEs it does — those cover several different characteristics, and the sphere-spacing error in particular is dominated by calibration scale, which is the kind of term the next paragraph shows averaging never touches.

Two catches, and they are the reason this report exists. The first is the independence condition in Step 4, which structured light does not satisfy. Projector defocus, phase error at fringe-period scale, meshing and normal estimation all correlate a point's error with its neighbours', so the effective N over a 20 mm patch is smaller than the point count by orders of magnitude — often enough to move the answer from a tenth of a micron to a figure in the low single-digit microns. Counting points is not the same as counting independent measurements. The second catch is that √N averaging only ever touches the random component at all. Calibration error, interreflection bias and smoothing filters are systematic: they survive any amount of averaging untouched, and they are exactly what turns 40,401 points into a confidently wrong plane. Point spacing tells you what you can resolve. It tells you nothing about what you can trust.

Which is the honest answer to "how accurate is a structured light scanner?" — it depends on the measuring volume, on the fit, and on which errors are random. Vendors say sub-millimetre; a calibrated sensor does far better than that on a fitted feature, as that acceptance certificate makes plain; and neither number bounds the uncertainty of measuring your casting's fillet radius, which is a task-specific uncertainty problem and a separate piece of work. The same caution applies to the fitting step itself: whether a plane or circle is fitted by least squares or by a minimum-zone criterion changes the answer, sometimes by more than the sensor noise — see Gaussian versus Chebyshev fitting.

3 · Why rough parts scan well — and what actually fails

The intuition most people bring — rough surface, bad measurement — is backwards for structured light. A shot-blasted or as-cast skin scatters the projected fringe pattern diffusely, which is precisely what a fringe-projection sensor wants. The failure cases on a casting are usually the machined patches: bright, specular faces that mirror the projector instead of scattering it, and dark glossy areas that swallow it. A Czech forging-machine builder, Šmeral Brno, bought a portable optical-tracking scanner — per Creaform's case study — specifically because it handled the shop's mix of glossy and rough surfaces with no preparation at all — something its laser tracker and CMM could not do. On the same logic, an iron foundry in Wisconsin — Willman Industries — inspects castings up to 30,000 lb with a sub-$20k handheld scanner and cut large-casting dimensional workups from 7–10 days of manual layout to 6–10 hours.

What actually corrupts scans of foundry parts is geometry, not texture. Deep pockets and concave features suffer multipath interreflection — inter-reflections, in much vendor documentation: projected light bounces between opposing walls before reaching the camera, adding false optical path length. The published analysis of the effect is worth knowing because the error is not what you'd guess — mild interreflection produces a systematic bias, but past a threshold the reconstruction develops a second, competing depth solution and phase unwrapping breaks down. The result is not noise you can average away; it is a confidently wrong surface. Practitioners on the ZEISS forum trade a concrete recipe for pockets: open the allowed sensor-to-surface angle up to 85°, accept overexposure on outer regions to get light into the pocket, use a smaller measuring volume, and give the rotary table more positions. Dark and shiny together — black glossy parts were the forum's example — is the combination practitioners describe as nearly impossible without matting. Transparent parts are worse still, and for the same reason: there is no opaque surface for the pattern to land on.

Interreflection belongs to a wider family the optics literature calls global illumination: every path by which projected light reaches the camera other than a single clean bounce off the point you meant to measure. Its other member matters in the foundry more than people expect. Subsurface scattering — light entering a translucent material, bouncing around inside it and leaving somewhere else — is exactly what an investment-casting wax pattern does, and to a lesser degree what resin-bonded sand cores and some polymer tooling do. The measured surface then sits slightly inside the real one, biased by a depth that depends on the material rather than the geometry — a different mechanism from interreflection, with the same character. Wax patterns are accordingly one of the clearest cases where coating a part before scanning is the correct call rather than a reflex, a case section 4's tolerance-band rule does not by itself cover. The general rule is worth stating once: structured light assumes each camera pixel sees light that bounced exactly once off an opaque surface, and every failure mode in this section is that one assumption breaking in a different way.

4 · The spray argument, with the numbers on the table

Which brings us to the longest-running argument in the field: matting spray. The debate has real numbers now, and they cut both ways.

A 2023 peer-reviewed study out of Brno UT measured eight commercial sublimating sprays with an ATOS III Triple Scan and found real coating thicknesses of roughly 24–43 µm at realistic multi-pass coverage — AESUB Blue at 32.6 ± 3.0 µm over four layers, Attblime AB Zero at 43.4 ± 9.2 µm. The manufacturer's datasheet for AESUB Blue says 8–15 µm per layer, and the study's per-layer figure agrees — the gap is that nobody sprays one perfect layer on a real part. Meanwhile the old-school answer, airbrushed titanium dioxide, measures around 3 µm — an order of magnitude thinner than any convenience spray (against ~44 µm for chalk spray) — it just doesn't vanish on its own. Sublimation times ranged from 47 minutes to 7.4 hours depending on product, over-spraying cracked some coatings, and one product left hundreds of holes in the mesh regardless of layers. The UK national lab's good-practice guide for optical point clouds says it plainly: expect matting sprays to cost precision "in the order of tens of microns."

The decision rule that falls out of the data: on a sand casting with millimetre-class tolerances, spray freely — 40 µm of coating is noise. On a precision casting or any machined datum you intend to measure optically at the 25–75 µm level, the spray is a significant fraction of your error budget: scan bare if the surface allows, switch to TiO₂ airbrushing if it doesn't, and treat "we always spray everything" as a process smell.

5 · The forging process, its defects, and which ones a scanner can see

Scanning discussions tend to assume everyone already knows what happened before the scan. Often they don't — and the list of forging defects is precisely the list of things a quality engineer is hoping the scanner will find. Some of them it finds beautifully. Some of them it cannot see at all, and buying an optical system expecting otherwise is how inspection plans get quietly wrong.

The four processes. The classification the Forging Industry Association uses splits the field four ways. Impression-die forging — closed-die — squeezes a heated billet between shaped dies, with excess metal escaping as flash around the parting line; this is the connecting-rod route and the process most of this report concerns. Open-die forging works the billet between flat or simply-shaped dies with no enclosing cavity, manipulating it between blows; it makes large shafts, blocks and rings, and its dimensional control is deliberately loose because generous machining stock is assumed. Seamless rolled ring forging pierces a billet into a doughnut and rolls it to grow the diameter while thinning the wall — bearing races, flanges, turbine cases. Cold forging forms at room temperature, buying tighter tolerances and a better surface at the cost of much higher tool loads. Two other classifications cut across that one, which is why the same part gets described three different ways by three different people in the same meeting: by temperature (hot, warm, cold) and by equipment (hammer, press, upsetter). The taxonomies overlap rather than nest — cold forging appears in the list above as a process in its own right and in the temperature axis as a class — so it is worth establishing which axis a supplier means before arguing about tolerances. The applicable tolerance standard is scoped by a combination of all three. EN 10243-1, the standard section 8 works from, covers hot steel die forgings produced on hammers and vertical presses; upset forgings made on horizontal machines fall under Part 2 instead.

The defects, sorted by whether light can reach them. This is the table worth pinning above the scanner.

Defect	What it is	Can structured light see it?
Underfill	Metal fails to fill the die cavity; the forging is short in a region	Yes — the classic full-field win. The colormap shows the starved region directly, with no hypothesis needed in advance.
Die mismatch (die shift)	Upper and lower die halves offset across the parting line	Yes — but it must be evaluated as its own characteristic per EN 10243-1, not folded into a profile result. See section 8.
Excess residual flash	Trimmed flash left proud of the die line	Yes — again a separate tolerance, not part of profile.
Warpage / bow	Distortion during cooling, trimming or heat treatment	Yes — and full-field is far better at it than point measurement, because the shape of the distortion is the diagnosis.
Progressive die wear	Cavity erodes over a die's life; the forging grows	Yes, over time — invisible in any one part, obvious in a trend of sampled scans. Section 6.
Laps and folds	Metal folded over itself with oxide trapped in the seam	Partly — a lap wide relative to point spacing shows as a groove; a tight lap narrower than the spacing is simply never sampled. Magnetic particle inspection (on ferromagnetic steels) or dye penetrant remains the detector.
Scale pits	Oxide pressed into the surface, leaving depressions	Partly — same sampling argument, and shallow pits fall inside surface noise.
Surface cracks	Cracks breaking the surface	Partly, and unreliably — width is usually far below point spacing. Not the instrument for this.
Internal cracks, inclusions, porosity, shrinkage	Discontinuities below the surface	No. Ultrasonic, radiographic, or industrial CT.
Decarburisation	Carbon depleted from the surface layer during heating	No — it has no geometry. Metallography or hardness traverse.
Core shift (castings)	Core displaced in the mould; walls thin on one side, thick on the other	No, not from outside. The textbook CT case: every external surface conforms while wall thickness is out of specification.

Read down the third column and the instrument's real shape appears: structured light is an outside-surface geometry sensor of exceptional quality, and the three-way split — owns it, partly covers it at the sampling limit, cannot see it — is a property of the physics rather than of any particular scanner. No amount of budget moves a defect from the third category to the first.

Where scanning sits in the inspection sequence. A common four-way split sorts inspection by when it happens: incoming or receiving inspection on purchased material, first article inspection to prove a new process, in-process inspection while parts are being made, and final inspection before shipment. The four are not a clean single axis — three are production stages while first article is event-triggered and can land at any of them — but the split is useful anyway, because full-field optical earns its place very unevenly across it. It is strongest at first article, where nobody yet knows which dimension will misbehave and measuring everything is exactly the right instinct — the argument made at length in the note on scan-based first article inspection. It is strong in-process as a sampled trend, which is the die-wear application. It is usually the wrong tool for high-rate final inspection, where a hard gauge answers one question in six seconds and the scanner answers a thousand questions in ten minutes.

6 · What foundries and forges actually do with scanners

Die and pattern correction is the founding use case. Scan the casting, compare full-field against nominal, correct the tooling — the deviation colormap replaced the layout table. Two published details matter more than the marketing version of this story. First, the reference isn't necessarily design CAD: Grede Foundries inspects patterns against a golden STL of the approved pattern, because tooling intentionally deviates from CAD to accommodate shrink — the colormap's zero is itself a negotiated artifact. Second, in the peer-reviewed die-correction workflow, registration uses machining datums rather than best-fit, because best-fit alignment absorbs exactly the systematic deformation the correction loop is trying to see. Grede's economics, from trade press: pattern-scan baselines eliminated three to five sample-casting iterations, and one brake-bracket job cut scrap from 5.2% to 1.0% — about $48,000 a year on that part alone. (Grede's instrument was a CMM-mounted laser scanner rather than structured light; the workflow is sensor-agnostic.)

Die-wear tracking without pulling the die: the Wrocław forging group publishes a technique they call reverse scanning — scan sampled forgings cyclically and watch the forging envelope grow over part count. The forging becomes the gauge for the tool, and die washout maps zone by zone while the die keeps running.

Scanning inside the forge itself is no longer exotic. Forges de Courcelles — reportedly Europe's second-largest crankshaft forger — runs automated ScanBox cells at three stations inside the crankshaft forging workshop, with press operators reading color deviation maps between strokes; per-part measurement that took 10–20 minutes now feeds SPC continuously (vendor-sourced case study). The physics of hot parts has been solved separately: a published in-line system measures forgings at ~900 °C with 0.11 mm repeatability on a 27-second cycle, using blue-light projection with a blue band-pass filter — blackbody glow concentrates in red and infrared, so the filter simply removes the incandescence — plus active cooling to keep the sensor alive. For everyone without that hardware, heat sets the practical rule: a conrod-sized forging leaves the die near 1,000 °C, cools initially at 100–150 °C/min — the rate falls away as the part approaches ambient, in the usual Newtonian fashion — and needs roughly ten minutes before thermal expansion stops dominating a 0.1–0.3 mm tolerance band. That published rule of thumb is a clock, not a criterion — and the arithmetic below shows which criterion it is standing in for.

Worked example — how much does a warm forging lie to you? (Hypothetical part, arithmetic shown.)

Setup. A drop-forged steel connecting rod with a nominal centre-to-centre distance of 150.00 mm, specified — as dimensional specifications are by default — at the ISO standard reference temperature of 20 °C. Its EN 10243-1 grade F centre-distance tolerance is ±0.4 mm (the figure from section 8). Take the coefficient of thermal expansion of plain carbon steel as α = 12.5 × 10⁻⁶ K⁻¹. That is a room-temperature value and it rises with temperature, so what follows is a first-order estimate that understates the error at high temperature — which is the safe direction for the argument being made.

Question 1 — scan it at 200 °C with no compensation. What does the scanner report?
ΔT = 200 − 20 = 180 K.
ΔL = L · α · ΔT = 150.00 mm × 12.5 × 10⁻⁶ K⁻¹ × 180 K.
First the product of the last two: 12.5 × 10⁻⁶ × 180 = 2.25 × 10⁻³.
Then: 150.00 × 2.25 × 10⁻³ = 0.3375 mm.
The scanner reports 150.34 mm for a part that is 150.00 mm at room temperature. That error is 0.3375 ÷ 0.4 = 84% of the tolerance limit, consumed before the forging process has contributed a single micron of its own variation.

Question 2 — how cool must it be for thermal error to fall to 10% of the limit (0.040 mm)?
ΔT = 0.040 mm ÷ (150.00 mm × 12.5 × 10⁻⁶ K⁻¹) = 0.040 ÷ (1.875 × 10⁻³) = 21.3 K.
So the part must be below 20 + 21.3 = 41 °C — near enough to room temperature to hold in your hand. For the record, the temperature at which thermal error equals the entire ±0.4 mm limit is ΔT = 0.4 ÷ (1.875 × 10⁻³) = 213 K, i.e. 233 °C.

Question 3 — don't wait; compensate. What does that leave? Scaling software corrects the measured geometry to 20 °C from a measured part temperature. If that temperature is known only to ±10 K — a fair estimate for a pyrometer reading one spot on a part with a thermal gradient through it — the residual error is
150.00 mm × 12.5 × 10⁻⁶ K⁻¹ × 10 K = 0.01875 mm ≈ 19 µm, or 4.7% of the tolerance limit.

Reading. Waiting and compensating buy the same thing, and the arithmetic prices both. Compensation is the cheaper one: at an identical part temperature it converts 338 µm of thermal bias into 19 µm, on nothing better than a ±10 K pyrometer reading. Waiting reaches the same neighbourhood only if you wait considerably longer — Question 2's 41 °C still leaves 40 µm on the table, and matching compensation's 19 µm means holding out for roughly 30 °C — and it pays in queue and work-in-progress rather than in error. On a hot line that is the expensive currency, which is why the instrumentation question is usually "where do I put the pyrometer", not "how long do I wait".

Two honest limits. First, this says nothing about how long ten minutes actually leaves a part at, because a cooling curve is exponential and the 100–150 °C/min figure is an initial rate — treat it as a linear rate for ten minutes and you get an answer below room temperature, which is obviously wrong. The published ten-minute guidance and this expansion arithmetic answer two different questions, and neither substitutes for measuring the part's actual temperature. Second, every number above assumes the part is one uniform temperature. A forging cooling from the outside in is not, and a scaling correction driven by one pyrometer spot cannot fix a thermal gradient. That residual is a genuinely hard problem rather than an arithmetic one.

Machining-stock verification closes the loop between foundry and machine shop: nest the finished-part CAD inside the scanned blank and confirm cleanup everywhere before the first chip. A steel foundry in one published account found 0.4–0.6 inches of unintended excess metal along a cope edge on first articles — a condition manual layout could not have mapped. And where the geometry is internal — cored passages, water jackets, porosity — structured light simply ends at the surface: that territory belongs to industrial CT, at $100k–$1M+ capex or a few hundred dollars per scan at a bureau, which is why it stays a sampled audit tool while optical covers dimensional work.

7 · Who leads, and why the software mattered as much as the sensors

The short version of the market: ZEISS owns the cell, Creaform owns the floor, and everyone else is fighting for third.

The ATOS line — blue-LED fringe projection with stereo cameras — came out of GOM of Braunschweig (founded 1990), which ZEISS acquired in 2019; in 2021 ZEISS also bought Capture 3D, its largest US distributor, so a North American foundry buying ATOS now deals with ZEISS-owned channel either way. The automated ScanBox cell debuted in 2012 with a foundry — Eisenwerk Brühl, cast-iron crankcases — among its first named installations, and the platform now spans series sized from 500 mm parts to full car bodies. The reference installations run from VW Kassel (Europe's largest light-metal foundry, ~200 components a day through a ScanBox plus tactile CMMs, reported through PiWeb) to Doncasters Bochum (investment-cast turbine blades, measuring times cut 2–3×) up to Siempelkamp, which measures sand castings to 320 tonnes in what ZEISS calls the world's largest non-contact robotic measuring cell (vendor-sourced case studies throughout).

Creaform — owned by AMETEK since 2013, which has since also absorbed FARO, consolidating the portable field under one roof — took the other regime: handheld and optically-tracked laser scanners with dynamic referencing, so part and scanner can both move on a vibrating shop floor. Named adopters include Fonderie Ariotti (structural castings, Creaform-reported 75 µm shop-floor accuracy), GF Casting Solutions, and — on the forge side — Šmeral Brno. The practical technology split, synthesized across every case I could find: fringe-projection cells win automated serial inspection; blue-laser portables win the floor, the very large, and the mixed shiny-rough surface without spray. Hexagon trails in structured light on foundry floors — my read of the case-study record — and answers with laser-based robotic cells; Artec's sub-$20k handhelds do real problem-solving work below metrology-cell grade. Budgetary classes, cross-checked across reseller listings: $25–60k entry metrology-grade handhelds, ~$80–100k+ configured portable systems, $150–300k+ for an ATOS-class system, $150–500k+ for automated cells — plus software that buyers routinely under-budget.

Those are sticker prices, and they are the wrong number to decide on. What matters is cost per part, which is a utilisation question. Take a hypothetical $200,000 cell over a five-year life running 2,000 parts a year: 5 × 2,000 = 10,000 parts, so $200,000 ÷ 10,000 = $20 per part in capital alone, or $20 × 2,000 = $40,000 a year. That sits just under the $48,000 a year one documented brake-bracket scrap reduction was worth — a useful sense of scale, though not an ROI, since that saving came from a different instrument on a different part. Run the same cell at 200 parts a year and the capital cost is $200 per part; run it at 20,000 and it is $2. Nothing about the scanner changed. Labour, fixturing, programming and software maintenance stack on top of all three figures, and for one-off service work they dominate completely — which is why the honest answer to "what does it cost to get something 3D scanned" is that the scan is rarely the expensive part, and published bureau rates vary too widely by part size, accuracy class and deliverable for a single number to mean anything.

One structural observation that rarely makes the brochures: the free viewer was a weapon. GOM Inspect's genuinely free tier meant a foundry could send a full inspection project — mesh, CAD comparison, GD&T, report — to any customer, who could open it at no cost. Deviation colormaps became the lingua franca of casting sign-off in large part because the reading software was free, and the file format lock-in pulled the hardware along behind it. Vendors understand ecosystems, not just optics.

8 · GD&T on parts that were never machined

Dimensioning an as-forged or as-cast part is its own discipline, codified in standards most machining-side engineers never open — and it is where scan-based inspection either respects the drawing's logic or quietly violates it.

Datum targets, not datum surfaces. A full as-cast face carries draft, parting-line edges, and gate grind-offs, so the drawing specifies where to touch: point, line, or area targets per ISO 5459 and ASME Y14.8, placed by theoretically exact dimensions. A DLA-sponsored metalcasting case study shows the craft: primary targets on a single mold-half's face, secondary and tertiary targets at identical elevation on 2°-drafted sidewalls so the draft cannot shift the part's located position — and the target map doubles as an instruction to the foundry about where not to put gating and parting features. In scan software, a 3-2-1 target scheme evaluated on the mesh is mathematically the same datum reference frame the fixture builder would have made with pins — if the software applies targets rather than best-fitting the whole surface.

Profile of a surface is the workhorse, because one callout can control a drafted, radiused, flowing as-forged shape against a target-based frame; ISO 8062-4 builds its entire general-tolerance scheme for castings on exactly that construction. The casting standards carry traps for the unwary evaluator: ISO 8062-3's general geometrical tolerances explicitly do not apply to features with draft — which is most of an as-cast surface — and wall thickness always takes one tolerance grade coarser than the part's general dimensional grade. For steel drop forgings, EN 10243-1 determines everything from five inputs (mass, die-line shape, steel category, shape complexity, dimension type), and its most operationally important rule is that mismatch, residual flash, and centre-to-centre tolerances apply independently of and in addition to all other tolerances. Software that folds die mismatch into a single profile evaluation is misapplying the standard — mismatch is its own characteristic, measured parallel to the die line at areas least affected by wear. One more contractual sharp edge from the same standard: the agreed forging drawing — not the customer's CAD, not the finished-part drawing — is "the only valid document for inspection of the forged part."

For scale, the numbers on a car-sized conrod forging under EN 10243-1 grade F: centre distance ±0.4 mm, mismatch 0.6 mm, residual flash 0.7 mm, each assessed separately. Hold those against the machined side of the same part: after fine boring, the big end is held to roundness in the tens of microns — close to an order of magnitude tighter than the as-forged figures above, and a form tolerance rather than a size or position one, so the comparison is of regimes rather than of like with like. That gap is why the rod needs more than one inspection regime rather than one good instrument.

9 · The connecting rod dossier

The rod at the top of this page is the report's case study because it compresses the whole industry into one part.

How they're made now. Drop-forged steel dominates Europe and China; North America reportedly runs mostly powder-forged rods (GKN's route, over half a billion installed since 1986). The big end is no longer sawn and machined as a separate cap: fracture splitting — notch the bore, shatter the big end in a press, bolt the halves back — has become the mainstream route, deleting up to half the machining steps. The metallurgy is deliberately perverse: C70S6 crack steel keeps sulfur high (0.045–0.07%) and yield low so the fracture runs clean and brittle; the stronger second-generation grades are actually harder to split well, fracturing slower with more tear-prone surfaces.

The unmeasurable feature. The fracture face itself carries no dimensional specification and must never be touched — the manufacturer's service literature forbids reworking it, resting the rod on it, even brushing it, because the joint's location accuracy comes from interlocking crack topography. An entire functional interface, deliberately outside metrology's reach: quality is inferred from proxies — big-end roundness after bolting (the split alone deforms a C70S6 big end by ~33 µm before the crack initiates, more than twice the second-generation steels), pairing-number integrity, and the fine-boring operation that follows.

Mass is a first-class characteristic. Rods carry (or carried) balance pads whose only function is to be ground away for weight matching — OEM sets historically matched to a few grams, race sets to ±0.5 g or better, weighed end-by-end in two-pan fixtures. A 1990s Opel machining line already did 100% in-line weighing with automatic classification at a ~6-second effective takt with one operator — weight grading is a sorting operation, not a lab measurement. And flashless precision forging closed the loop from the other side: billet volume control got good enough that the small-end balancing boss — and its machining operation — could be deleted entirely.

Where scanning enters. In conventional hammer forging, flash consumes an astonishing share of the billet — a recent process study measured 61% of the charge going to flash, cut to 49% by a redesigned preform, with the improvement verified by 3D scan colormaps every hundred pieces. That's the pattern across the published record: structured light lives at the forge — die fill, die wear, mismatch, preform development, sampled full-field conformance after the ten-minute cooldown — while the machining line stays hard-gauge (air gauges and bore gauges at 6-second takt, roundness checked with the cap torqued) and assembly owns weighing and crack proxies. Three inspection regimes, one part; the scanner owns exactly one of them, and that's not a limitation — it's the correct division of labor.

The market context, for anyone planning QC investment: hybrids, not battery EVs, set the conrod demand curve — every HEV and PHEV still contains rods — and one market forecast (vendor forecast, treat accordingly) has the market growing through 2036 on "demand redistribution rather than disappearance." The observable engineering response is automation of legacy forge lines rather than new capacity: recent papers study robotizing existing hammer lines at plants running millions of rods a year. Defending margin on mature volume is precisely the environment where scrap-rate and die-life economics — the things forge-side scanning measures — pay for instrumentation.

10 · Can you trust the scanner? Standards and the acceptance argument

The argument between "the scanner is certified" and "my customer only accepts CMM data" has a real technical substrate, and 2021 changed its terms.

For two decades the acceptance framework was a German guideline, VDI/VDE 2634 — probing error on a sphere, sphere-spacing error on a ball bar, flatness on a plane, sized to the measuring volume — the exact set depending on which sheet applies. Both relevant sheets are now formally withdrawn in favor of ISO 10360-13:2021, the first ISO acceptance standard written for self-contained optical 3D scanners — yet vendor certificates issued as recently as 2025 still say "with reference to VDI/VDE 2634 Part 3," the paperwork lagging the standards landscape by years. A real 2025 acceptance certificate for a small ATOS measuring volume shows what the numbers look like in practice: MPEs of 3–13 µm depending on characteristic, DAkkS-traceable sphere-bar artefact, measured results roughly 3–10× inside the limits.

ISO 10360-13 is a better test — six characteristics instead of three, including distortion and concatenated-volume length errors — and it contains one clause every buyer should know: results are evaluated at both a 95th-percentile and an all-points basis, explicitly because the difference "can reveal influences of smoothing filters… not always transparently visible for users." A standard written to police spec-sheet games. Two hard caveats keep it honest. NIST showed in 2024 that even the improved length tests fail to consistently detect all systematic calibration errors in structured-light systems — a passed certificate is weaker evidence than most quality engineers assume. And the standard explicitly excludes handheld laser-line scanners — arguably the most common portable instruments on foundry floors — leaving them without a dedicated ISO acceptance route at all.

On the correlation question, the measured picture is bias, not noise: against a tactile CMM, structured-light step heights in a DOE-funded study read 23–36 µm high with non-overlapping error bars, and small-hole diameters erred by up to ~125 µm, improving rapidly with hole size. The same study demolished the standard gauge R&R framing: a 2.9 mm hole's diameter spread collapsed from ~65 µm to ~8 µm just by scanning from 15 positions instead of 5 — the "operator" in an optical MSA is really scan-position count and meshing, which classical AIAG crossed studies were never designed to isolate. Hence the doctrine you actually see in regulated supply chains: hybrid inspection — tactile CMM holds the datums and tight primitives, full-field optical owns freeform surfaces and wall thickness, X-ray or CT owns the inside — and OEM acceptance of scan data turns out to be procedural, not physical: in the Boeing chain it's digital-product-definition process approval (measurement software verification included), and AS9102 Rev C explicitly contemplates verification against 3D model data. The scanner isn't on trial; your process documentation is.

11 · What this means if you run the QC lab

Point spacing is not accuracy: spacing is fixed by measuring volume divided by camera resolution; a fitted feature averages thousands of points and does far better. But the points are not independent to begin with, so effective N sits far below the point count — and √N averaging only ever kills random error anyway; calibration bias, interreflection and smoothing survive it intact.
Surface logic beats instinct: as-cast and shot-blast surfaces are friendly targets; machined datums, dark glossy patches, and deep pockets are where scans fail — and interreflection fails confidently, not noisily. Translucent wax patterns fail the same way, biased inward by subsurface scattering.
Know what the instrument cannot see: optical scanning owns shape defects — underfill, mismatch, flash, warpage, die wear. It is blind to porosity, inclusions, decarburisation and core shift. A quality plan that lists a scanner against an internal requirement has a hole in it.
Temperature is a first-order error, not a footnote: on a 150 mm steel dimension, every 100 K above 20 °C is ~0.19 mm. Compensating from a ±10 K reading leaves the same residual error as waiting for the part to reach about 30 °C, at a fraction of the queue time; both beat ignoring it.
Spray by the numbers: convenience sprays measure 24–43 µm at real coverage; TiO₂ ~3 µm; bare where possible. Match the decision to the tolerance band, not to habit.
Registration is a policy decision: best-fit hides exactly what die-correction loops need to see; datum-target evaluation on the mesh must replicate the drawing's target scheme. Mismatch and flash are separate characteristics by standard, not components of profile.
Know which regime you're in: forge-side sampled full-field, machining-side 100% hard gauge, assembly-side proxies. Structured light owning one regime is the design, not the deficiency.
Acceptance ≠ task uncertainty: a sphere-bar certificate — VDI 2634 or ISO 10360-13 — bounds nothing about your casting's fillets. Task-specific uncertainty needs substitution-style studies on calibrated parts, and optical MSA must vary scan strategy, not just operators.
The deliverable question is real: golden mesh, shrink-compensated tooling model, or design CAD as reference changes what a colormap means. Decide it in writing before the first scan.

Everything in that list is pipeline work — the evaluate-decide-move-deliver-defend chain that starts where the scanner stops. If your foundry or forge shop has the sensor and the colormaps but the decisions above were made by defaults, that is exactly the work I offer — starting with a fixed-fee pipeline assessment: hello [at] metrologymaven [dot] io.