INDUSTRY REPORT 03 · INDUSTRIAL CT

Industrial CT: the inspection that sees what line-of-sight never will

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

A closed impeller carries its most consequential surfaces where no probe stylus, no camera, and no fringe projector will ever reach: the flow passages sealed between hub and shroud. For most of manufacturing history the honest options were to cut the part apart or to trust the process. Industrial CT is the third option — the one inspection technology whose reach does not depend on line of sight — and it has quietly become the reference method for castings and additive parts whose critical geometry is internal. This report covers what computed tomography actually produces, what it is trusted for today, where its numbers go soft, and why the labs that own a scanner drown in evaluation time rather than scan time.

1 · The closed-impeller problem — and what a CT scan actually produces

Every other instrument in the metrology lab is a line-of-sight instrument. A CMM stylus must physically reach the feature; a structured-light sensor must see it from two camera angles at once; even a borescope only inspects the corridor it crawls through. The impeller at the top of this page defeats all of them by design: the passages between the vanes are bounded above by the shroud, and there is no direction from which a probe or a projected fringe enters. The same is true of cored oil galleries in a cylinder head, conformal cooling channels in an additively built mold insert, and the sealed interior of any assembled product. CT scanning does not care, because X-rays pass through the part, and the measurement is built from what the part did to them on the way. That is what industrial computed tomography is: a non-destructive testing method that recovers a part's internal structures from X-ray transmission rather than from any view of a surface.

It helps to be precise about what the machine actually produces, because it is not a point cloud. Inside a cabinet CT scanner — an arrangement that section 2 shows is usual but not universal — the geometry is fixed and the part moves: an X-ray source on one side, a flat-panel detector on the other, and the scanned object on a rotary stage between them. The object rotates through a full turn while the detector records hundreds or thousands of two-dimensional projections — greyscale radiographs, each one a map of how much the beam was attenuated along every ray through the part. Those projections are the raw data, and they are worth naming separately from the CT images a technician later scrolls through, which are cross-sectional slices cut back out of the reconstructed volume. A reconstruction algorithm (for the standard cone-beam geometry, the Feldkamp method and its descendants) folds those projections into a three-dimensional grid of grey values. Each cell of that grid is a voxel — a volumetric pixel — and its grey value encodes the local X-ray attenuation, which tracks density and material. That volume is the raw product of computed tomography. Note what it is not: it is not yet a surface, not yet a dimension, not yet a defect count. It is a block of numbers in which material is bright and air is dark and the boundary between them is a gradient several voxels wide.

Surface determination is the step that turns grey values into a measurable boundary, and it deserves more respect than it usually gets. The classic starting point is the ISO-50 rule: take the grey value halfway between the material peak and the background peak of the histogram, and call that iso-surface the part. Modern software refines this locally — adapting the threshold region by region to the actual gradient — and interpolation across the gradient locates the boundary to a small fraction of a voxel, which is why CT can measure features finer than its voxel size suggests. But here is the part that matters for anyone signing an inspection report: every downstream number inherits the surface-determination decision. Shift the threshold and every hole in the part grows while every pin shrinks, wall thicknesses change in lockstep, and marginal pores swell or vanish. Two analysts with the same volume and different surface settings will produce two different parts. That makes surface determination a process parameter to be chosen, validated, and written down — not a default to be accepted. Once the surface exists, the familiar coordinate-metrology machinery takes over — alignment, feature fitting, GD&T — and the same least-squares-versus-Chebyshev choices that govern any fitted feature apply to CT-extracted geometry too.

One more piece of vocabulary discipline. Voxel size is set by geometry: the closer the part sits to the X-ray source relative to the detector, the higher the magnification and the finer the voxels — so small parts scan fine and large parts scan coarse, and a claim like “5 µm accuracy” with no part size attached is marketing, not metrology. And voxel size is not resolution: focal-spot blur, detector characteristics, and artifacts all limit what the volume genuinely resolves, which is why the German guideline family for dimensional CT (more on it in section 5) treats structural resolution as its own characteristic, separate from voxel count. High-resolution micro-CT of a fingertip-sized sample genuinely reaches voxel sizes of a few micrometres; the same class of industrial CT scanners pointed at a cylinder head resolves orders of magnitude coarser, and the spatial resolution actually achieved depends on focal-spot size, geometry, and the image quality of the projections — not on the voxel number in a file header. A defect that spans one or two voxels may well be real; its measured size — two opposing edges inside a single blur kernel, rather than one boundary interpolated across it — is a rumor.

2 · Industrial CT versus medical CT: same mathematics, opposite priorities

Almost everyone arrives at this subject through a hospital, and the search results for it are still half medical. The confusion is worth clearing early, because the two machines share their physics and their reconstruction mathematics and agree on almost nothing else. Both are X-ray computed tomography. Both recover a three-dimensional volume from a set of transmission projections. Then their constraints diverge, and every visible difference follows from one fact: a patient must not be irradiated heavily and must not be asked to hold still for an hour; a casting does not care about either.

That single asymmetry produces the rest. In medical imaging the patient lies still and the gantry — source and detector together — spins around them, because rotating the subject is not an option; the scan is over in seconds, the dose is held to the minimum that yields a usable image, and the tube is tuned for contrast between soft tissues rather than for penetrating metal. In the cabinet systems that make up most of the industrial installed base the arrangement inverts: the source and detector stand still and the object rotates on a stage. The scan may then run for minutes or hours, the dose is limited by the tube and the cabinet's shielding rather than by the subject — at least when the subject is metal — and the voltage is chosen to get signal through aluminum, titanium, or steel. (Rotating-gantry and robot-guided industrial systems do exist, for objects too large or too fixed to spin; they are built precisely where the object cannot be the thing that moves.) The consequences for anyone comparing specifications:

Two practical points follow. The first is a purchasing trap: medical scanners are specified and priced in slices — roughly, the number of cross-sections acquired per rotation — and price lists quoted per slice count are medical vocabulary that does not describe an industrial system at all. An industrial CT scanner is specified by tube voltage and power, focal-spot size, detector format, achievable voxel size at a given part envelope, and stage accuracy. If a quotation for industrial work is denominated in slices, someone has cut and pasted from the wrong market. The second is not a deduction from the table but a correction to the assumption behind it, and it is worth stating plainly: the traffic runs both ways, because medical device manufacturers are among the heaviest users of industrial CT, scanning catheters, implants, and sealed drug-delivery assemblies on exactly the machines described in this report.

And the MRI question, which follows the medical one so reliably that it deserves a sentence. Magnetic resonance imaging is not a variant of CT; the two are both tomographic, and there the resemblance ends. MRI has no X-ray source at all — it excites hydrogen nuclei in a strong magnetic field and reads the signal they emit as they relax, which makes it excellent at distinguishing water-rich soft tissues and essentially useless on a metal casting, which contains no mobile hydrogen to excite — and which, if it is ferrous, has no business near a superconducting magnet in the first place. In dimensional engineering there is no MRI-versus-CT decision to make. The comparison that matters is CT against the methods in the rest of this report.

One last import from the medical side, which searchers keep asking and which needs translating rather than answering: is there anything you shouldn't do before a CT scan? The patient-preparation advice behind that question — fasting, contrast agents, removing jewelry — has no industrial counterpart, because the subject is a part. The industrial version of the question is real, though, and it has four answers. Get the part clean and dry: trapped coolant and oil are water-dense and will read as material. Fixture it in low-density foam or plastic rather than in a steel vice, because dense fixturing starves the beam and streaks the volume. Orient it so that the longest metal path does not run through the features you care about. And let it reach room temperature, because a warm part drifts over a long scan. None of these are exotic. All of them are the difference between a volume you can measure and a volume you can only look at.

3 · What industrial CT is trusted for today

Four industrial applications carry most of the installed base, and they have one thing in common: in each of them the quality control decision depends on something the part will not show you from outside.

Porosity and inclusion analysis is the flagship, especially in aluminum die casting. Gas porosity, shrinkage porosity, and foreign-material inclusions are internal defects by definition — the whole failure mode lives below the surface — and CT is the only method that maps them in 3D without destroying the part. It is worth being clear about the mechanism, because it is what makes internal defect detection possible at all: the reconstruction encodes local material density, so a void reads as a region of low attenuation and a tungsten inclusion as a bright one, and flaw detection is a matter of finding density that does not belong rather than of seeing a surface. Mature porosity analysis does not ask “is there porosity?” (the answer in a die casting is essentially always yes); it asks whether the pores exceed an agreed acceptance class: limits on pore size, local concentration, and — critically — distance to a machined surface, because a pore that the finishing cut will open into a sealing face is a leak path, while the same pore buried mid-wall may be irrelevant. In German-speaking foundry practice the anchors for this are the foundry association's reference-sheet family — P 201 and P 202 on volume deficits, with P 203 extending the evaluation methodology to CT specifically. I am confident of the family's role in drawing callouts; check the current editions before writing one into a purchase spec. The classifiers also use pore shape: compact, roundish voids read as gas porosity, sprawling dendritic ones as shrinkage — different root causes, different corrective actions at the machine.

Wall thickness on complex castings is the second pillar, and the turbo housing above shows why. A cored casting's wall thickness is the distance between an outer surface you can see and an inner surface you cannot; if the core shifted during pouring, the wall is thin on one side and thick on the other, and nothing visible from outside betrays it. CT wall-thickness analysis computes the local thickness over the entire part and paints it as a colormap — full-field, including every cored passage — which turns core shift from an occasional sectioning discovery into a routinely monitored characteristic. For thermally and pressure-loaded parts like turbine and turbocharger housings, minimum wall is often the life-limiting dimension, and CT is the only instrument that measures it everywhere.

Assembly inspection exploits the same physics in the other direction: scan the finished, assembled product and inspect it in its working state — seated seals, engaged clips, solder joints, wire routings — without taking it apart, which for many failure modes is the only meaningful time to look. Assembly verification of this kind is largely a question about component placement: is the connector fully seated, is the clip engaged, did the potting compound reach the corner. On electronic assemblies the same scan resolves solder voiding and bond-wire routing inside a package that cannot be opened without destroying it. A documented example from my research corpus: ZKW, the lighting-systems supplier, brought CT in-house with a METROTOM and ZEISS INSPECT X-Ray to inspect assembled headlights in batches of more than twenty, after external CT services had meant waits of three to four weeks (vendor-sourced case study). Multi-material assemblies — plastics, metals, electronics in one scan — are exactly the case where sectioning destroys the evidence it is looking for.

Additive manufacturing inspection may be the application that made dimensional CT mainstream, because AM removed the alternative. The entire argument for printing a part is often its internal geometry — conformal cooling channels, internal lattices, consolidated flow passages — and none of it is inspectable by any line-of-sight method after the build. CT verifies channel cross-sections, finds trapped powder, and measures lattice struts; independent researchers at Brno UT/CEITEC, for instance, scripted the creation of inspection elements across the many struts of laser-melted lattice structures measured by micro-CT — a task nobody could do by hand at scale (independent paper, via research corpus). On the standards side, ASTM E3166 catalogs nondestructive examination methods for metal AM aerospace parts, CT prominent among them. The honest caveat: the metals AM favors vary in X-ray friendliness — aluminum and titanium scan well at moderate energies, dense nickel superalloys demand more voltage and yield noisier volumes.

A pattern worth naming across all four: the operations that trust CT most got there by correlation, not faith. The die caster TCG UNITECH, which replaced several separate inspections with one CT scan per part, reportedly spent months correlating CT porosity results against destructive microscopy and other methods before relying on them (vendor-sourced case study). That is not a weakness of the method. That is what adopting any new reference method correctly looks like.

4 · Failure analysis, reverse engineering, and first article inspection

The four pillars above are production work — the same part, over and over, against an agreed rule set. Three further uses are worth separating out, because they are one-off investigations rather than recurring inspections, and the discipline each one demands is different.

Failure analysis is the case where CT's non-destructive character stops being a convenience and becomes the whole argument. A part that failed in service is evidence, and it is the one part you cannot afford to cut, because the first saw stroke destroys the geometry you are trying to explain and cannot be repeated. Scanning first inverts the risk: the volume is a permanent, re-openable record of the part as it arrived, so sectioning — when it happens — is aimed rather than exploratory, and the analyst can go back to the volume six months later when a second failure arrives and the question has changed. What CT finds in this mode is the usual catalog of things that compromise structural integrity from inside: fatigue cracks propagating from a subsurface pore, lack-of-fusion in a weld, delamination between composite plies, a fastener that never seated. The caution here refines what section 1 says about voxels, and the refinement is the difference between finding a defect and sizing one. A crack narrower than a voxel does not appear as a thin crack; it appears, if at all, as a slight depression in grey value spread across the voxels it crosses, and whether that survives the noise depends on its width, its orientation to the beam, and the contrast available. Favorably oriented sub-voxel cracks are detected that way every day; unfavorably oriented ones are missed. But detected is all they are — section 1's warning that the size of a one- or two-voxel feature is a rumor rather than a measurement applies with full force the moment anyone tries to put a width on such a crack. Report it as found, not as measured, and state the detection limit rather than leaving the reader to assume one.

Reverse engineering is CT used to produce geometry rather than to judge it: surface-determine the volume, extract a mesh, and fit CAD data to it. This is genuinely valuable where no drawing survives — a legacy pump housing, an obsolete casting whose tooling is gone, a competitor's part in a teardown — and it is the one application where CT's simultaneous access to internal and external features is decisive, because a scanned-and-fitted model of the outside alone is a shell, not a part. Two honest cautions belong with it. The first is that a reverse-engineered model inherits every surface-determination decision described in section 1, so two analysts will hand you two slightly different CAD models of the same object. The second is more often forgotten: a model built from one part records that part, not the design. The as-cast draft, the wear, the shrinkage, the local core shift — all of it is faithfully captured as though it were intent. Reverse engineering produces a starting point for an engineer, not a drawing to manufacture from.

First article inspection is where the two worlds meet, and it is a natural fit for exactly one reason: an FAI must verify every design characteristic, and on a cored casting or an additively built part some of those characteristics are internal. AS9102 does not name computed tomography. An FAI has to account for every design characteristic — the requirement, the result, and how it was verified — which means CT enters an FAI package the way any other method does: as a method the customer accepts for that characteristic, not as a brand name. The practical value is coverage: one scan carries wall thicknesses, internal passage dimensions, and porosity evidence that would otherwise need sectioning of a sacrificial part, which for a first article is an expensive way to prove the first article was good. Where the external characteristics matter more than the internal ones, the trade-offs are the ones set out in the FAI report, and they mostly favor optical scanning or a CMM. Match the method to where the characteristic lives.

5 · The metrological honesty section: standards, artifacts, uncertainty

CT dimensional measurement is real — parts are accepted and rejected on it every day — but anyone selling it as a solved problem is skipping the interesting part. The acceptance-testing framework for CT is younger and messier than the tactile world's, and the physics contributes biases that a CMM user never has to think about.

The standards situation. Tactile CMMs have had ISO 10360 acceptance and reverification tests for decades. For CT, the working framework since roughly 2010 has been the German guideline family VDI/VDE 2630 — Part 1.2 on the influence quantities, Part 1.3 on applying ISO 10360-style tests to CT sensors, Part 2.1 on determining task-specific measurement uncertainty. An actual ISO acceptance standard, ISO 10360-11 for coordinate measuring systems using X-ray CT, arrived only in 2023. If that trajectory sounds familiar, it is the same story I told about structured light in report 01 — a German guideline carrying the field for a decade until an ISO part lands — except CT's ISO part is two years younger still, and the paperwork in circulation will reference the VDI framework for years yet. Meanwhile the NDT community has its own, older CT standards — ASTM E1441 (guide) and E1570 (practice) for CT examination, ASTM E1695 for system performance, and the ISO 15708 series on radiographic CT methods — which govern defect detection, not dimensional acceptance. That literature grew out of industrial X-ray radiography and treats CT as its three-dimensional extension, so it speaks the language of flaw sizing and probability of detection rather than of measurement uncertainty. Two communities, two literatures, one machine: a lab that quotes a non-destructive testing standard as evidence of dimensional capability, or vice versa, is waving the wrong certificate. The distinction also matters when a customer asks whether CT can replace destructive testing outright — for finding and sizing internal flaws it very often can, and that is the older ASTM literature's home ground; for proving the dimensional accuracy of a characteristic to a stated uncertainty it can only do so under the VDI/ISO framework above.

The artifacts. CT's systematic errors have names, and they bias numbers rather than merely blurring pictures. X-rays are attenuated to varying degrees depending on material density and thickness, and the first two artifacts below are what happens when reconstruction's assumptions about that attenuation do not hold. The table gives the plain-language versions; all of this is standard CT physics, stated as general knowledge.

The practical consequence of the first two rows is worth underlining: beam hardening and scatter do not add random noise you can average away — they move surfaces, systematically, by amounts that depend on material, geometry, and orientation. This is why a CT machine's sphere-distance acceptance test — typically run on a friendly calibrated artifact — bounds very little about your actual casting, and why task-specific uncertainty is the concept that matters. The honest route to a defensible number is the substitution approach codified in VDI/VDE 2630 Part 2.1 (the CT sibling of the ISO 15530-3 method): measure a calibrated twin of your part, on your machine, with your exact scan recipe and evaluation template, and let the observed errors set the uncertainty. It also reframes measurement-systems analysis: in a CT gage study, the “operator” barely matters compared to the parameter set — scan recipe, surface determination, evaluation template. A classical crossed R&R that varies people while holding parameters fixed will flatter a CT process; vary the things that actually vary.

A newer entry belongs in the same section, for the same reason. Artificial intelligence has arrived in CT image processing — machine-learned denoising and reconstruction now ship in commercial software, and they do what is claimed for them, which is to produce a cleaner-looking volume from fewer projections or a shorter exposure. Treat the claim precisely, though: those methods improve image quality, and image quality is not the same quantity as measurement traceability. A denoiser trained on other parts has a prior about what surfaces ought to look like, and a prior that fills a marginal pore or straightens a noisy edge has changed the measurand. The safe posture is the one the rest of this section argues for — keep the raw data, and validate the whole chain, learned steps included, against a calibrated twin before any of it is allowed near an acceptance decision.

6 · What an industrial CT capability actually costs

The question everyone asks first is what the machine costs, so here is the honest shape of the answer, followed by the reason it is the wrong question to stop at. Capital is serious money: industrial CT systems run from roughly $100k at the entry end to well past $1M for high-energy or large-cabinet configurations (vendor-sourced cost classes, consistent with report 01). What moves a system along that range is not a single headline number but a stack of them. Tube voltage and power come first, because penetrating steel is what costs money. Then focal-spot size, which sets how fine a voxel the geometry can deliver; detector format and quality; and the size and shielding of the cabinet, which scales with both the part envelope and the energy. Rotary-stage accuracy is a metrology cost rather than an imaging one, and it is easy to under-buy because little in the picture reveals it at the magnitudes that matter dimensionally. Last comes the evaluation software, frequently licensed separately, and it is where a surprising share of the capability actually lives.

Two warnings about the numbers in circulation. The first is the slice-count trap from section 2: figures quoted per slice — a 16-slice system at one price, a 256-slice at another — describe medical scanners, and pasting them into an industrial budget produces a number that means nothing. Industrial CT scanners are not sold by slice count. The second is that used and refurbished systems are a real market and do trade at a substantial discount, but the discount depends so heavily on tube hours, detector condition, and whether the software licenses transfer that a percentage quoted in the abstract is not information. Get the tube hours in writing.

Before any of that, though, there are industrial CT scanning services. A service bureau will scan and evaluate a part for a few hundred dollars, and for most organizations that is not merely the cheap option but the correct first one: it converts a capital decision into a per-part cost while you find out whether CT actually answers your questions, and it produces the correlation evidence section 3 argues you need before trusting the method anyway. The failure mode of the bureau route is turnaround rather than price, and it is the reason organizations eventually bring scanning in-house — the external lead times behind ZKW's decision in section 3 are survivable for a failure investigation and fatal for a production disposition. The decision rule is close to arithmetic: bureaus win while scans are occasional and the answers can wait; in-house wins when the queue becomes routine or the part cannot leave the site.

And then there are the costs that do not appear on the quotation. A shielded enclosure has floor-loading and siting requirements. Radiation safety brings a license, a survey, and someone's ongoing responsibility for it. Service contracts and periodic calibration are annual, not optional. CT data is large — a single volume runs to gigabytes, so a year of production scanning is a storage and retention decision, not a folder. None of these is the big one. The big one is the last line of the quotation nobody writes down, and it has its own section.

7 · Throughput economics: the hours are in evaluation, not the scan

The economics of industrial CT changed shape over the last decade, and the shape it took surprises buyers. Scanning got fast — small-part scans that once took an hour run in minutes, and reconstruction that once tied up a workstation overnight finishes on a GPU before the next part is fixtured. Capital, as section 6 sets out, is a known and budgeted quantity. But the line item that actually swells as a CT program succeeds is neither of these. It is analyst evaluation time.

Consider what happens after every scan of a recurring part: someone opens a multi-gigabyte volume, runs surface determination, aligns to the datum structure, applies the porosity rule set, checks wall-thickness zones, evaluates the dimensional characteristics, and assembles a report. Done by hand, that is tens of minutes to hours of skilled attention per part — repeated identically for part after part. The pattern in my own buyer research is blunt: the CT teams feeling the most pain are precisely the ones with recurring multi-part or defect workflows — large data, repeated segmentation and evaluation, high analyst hours, and complex exports into quality systems. The scanner keeps up; the human doesn't. ZKW's batches of twenty-plus assemblies per run only work because the evaluation is templated, and Festo — facing CT reproducibility across a portfolio it counts at 33,000 products and thousands of part variants — distributes a common evaluation plan globally and uses an automated defect-detection application to classify porosity (vendor-sourced case studies).

It is worth putting arithmetic to that, because the number is larger than most buyers expect and it is the number that justifies the software rather than the scanner. The following figures are illustrative and hypothetical — they are not measurements from any client, and the point is the structure of the calculation, not the inputs. Take a lab running 600 CT inspections a year on one recurring part family — about a dozen a week. Suppose hand evaluation takes 45 minutes per part and a validated template reduces that to 8 minutes of loading, running, and reviewing exceptions. Cost the analyst at a fully burdened $75 per hour, and suppose building and validating the template takes 60 hours once.

Three ways to the same figure, because a number nobody checks is decoration. By hours: 600 × 45 min = 450 h by hand against 600 × 8 min = 80 h templated, a difference of 370 h, and 370 × $75 = $27,750. By minutes saved: 600 × (45 − 8) = 22,200 minutes = 370 hours, the same 370. By cost per part: $56.25 − $10.00 = $46.25 saved on each, and $46.25 × 600 = $27,750. Subtract the one-time $4,500 and the first year nets $23,250; the template pays for itself after roughly 98 parts, or about two months at this volume. Note what the calculation does not claim — nothing here says the templated answer is a better answer. It says that once the template exists, the same answer costs less than a fifth as much to produce — and that is only worth having if the answer was right to begin with.

This is where evaluation automation earns its keep, and it comes in escalating grades: parametric project templates that re-execute an entire evaluation on the next volume; batch processing that runs the template across a directory of scans overnight; automated defect rules scoped by region of interest so the software flags only what the acceptance class actually cares about; and scripted pipelines that push results straight into the quality system. The tooling has been moving the same direction — the 2026 release of ZEISS INSPECT X-Ray, to take the platform I work in, expanded its Python scripting API and Q-DAS statistical export alongside multi-material and defect-analysis improvements (release notes, via research corpus). One honest caution belongs in the same paragraph: automation does not make a measurement more true. It makes the same decisions consistently, at scale, and moves the analyst's hours from repetition to exceptions — which is exactly what you want, provided the templated decisions were validated first. Automating an unvalidated surface-determination setting just manufactures wrong answers faster. The scripted-evaluation work I show on the examples page is this category of work: the pipeline is the deliverable, and the validation evidence rides with it.

8 · CT and structured light: inside and outside

Readers of report 01 will remember the division of labor that emerged from the foundry evidence: tactile CMMs hold the datums and tight primitives, full-field optical owns the freeform outside, and CT owns the inside. Nothing in this report revises that doctrine — it fills in CT's side of it. The two full-field technologies are complements, not competitors, and the boundary between them is nearly clean: structured light is faster and cheaper per part on everything it can see, and it can see nothing it doesn't have line of sight to; CT sees everything and pays for it in cycle time, capital, and artifact management.

The genuinely interesting question is the overlap case: CT reconstructs the outside of the part too, so a single volume holds the internal and external features together, and a part that needs CT anyway — for porosity, say — can in principle carry its external dimensional evaluation in the same scan and skip a second measurement. Measuring internal and external structures at once, in one alignment, is a genuine advantage and no other method offers it; single-scan combined workflows are a real selling point of modern CT software. But it should be a decision, not a drift: external features measured through CT carry the surface-determination and artifact budget of section 5, and on dense or awkwardly proportioned parts an optical scan of the exterior will beat the CT-derived exterior on both uncertainty and cost. Physics draws the outer boundary of the whole method: aluminum, magnesium, plastics, and small steel parts are CT-friendly; large dense steel sections starve the beam, which is why the big steel castings of report 01 get optical outsides and sampled, high-energy — or destructive — insides. Match the modality to the material and the tolerance, one characteristic at a time.

9 · What to decide in writing

Every recurring theme in this report converges on the same discipline: the decisions that determine what a CT number means are all upstream of the number, and they should exist on paper before the first acceptance decision. The list is short.

• Surface determination method — which algorithm, which settings, per part family; and the validation evidence (correlation to tactile or destructive reference) that justified it.
• Scan recipe — voltage, filtration, magnification, and therefore voxel size, frozen per part family the way a CNC program is frozen.
• Porosity acceptance class — the named rule set: pore-size limits, concentration limits, distance-to-machined-surface rules, minimum defect size honestly stated in voxels, and who owns borderline calls.
• Correlation study — the reference-method comparison, done before trust, TCG-UNITECH-style, and retained as evidence.
• Evaluation template ownership — who may edit the template, how changes are versioned and re-validated; an unversioned template is an uncontrolled gage.
• Data retention — volumes are gigabytes each; decide what is kept, for how long, and what merely the report must prove.

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

Sources

Dimensional-CT standards: ISO 10360-11:2023 (acceptance/reverification tests, CT-based coordinate measuring systems) · VDI/VDE 2630 guideline family — Blatt 1.2 (influence quantities), Blatt 1.3 (ISO 10360-style testing for CT sensors), Blatt 2.1 (task-specific measurement uncertainty); VDI, Düsseldorf · ISO 15530-3 (uncertainty by calibrated workpieces — the substitution method Blatt 2.1 parallels)
NDT-side standards: ASTM E1441 (Standard Guide for Computed Tomography) · ASTM E1570 (Standard Practice for CT Examination) · ASTM E1695 (CT system performance measurement) · ISO 15708 series (radiation methods for computed tomography) · ASTM E3166 (NDE of metal additively manufactured aerospace parts)
Adjacent regimes: AS9102 (aerospace first article inspection requirement) — cited in section 4 for what an FAI accounts for, not as a CT standard; it does not name computed tomography
Porosity acceptance: BDG (German foundry association) reference sheets P 201/P 202 (volume deficits of castings) and P 203 (porosity analysis by CT) — cited for their role; consult current editions before contractual use
Vendor case studies & corpus: ZEISS success stories: ZKW (assembled-headlight CT, external lead times), TCG UNITECH (CT correlation before trust) · ZEISS / Festo (automated defect classification at portfolio scale) · ZEISS: CT cost classes · ZEISS INSPECT X-Ray Release 2026 notes (Python API, Q-DAS, defect-analysis expansion) · Brno UT/CEITEC micro-CT lattice automation study (independent paper) · practice research corpus on ZEISS INSPECT workflows and buyers, 2026 (unpublished; basis for the “evaluation is the bottleneck” pattern)