Industries / Mining and materials / Steel and metals / Foundries and casting
Inspection
Inspection is checking finished castings for hidden flaws, such as gas holes, shrinkage cavities, trapped sand and cracks, mostly with X-rays and without cutting the part open.
5
Milestones achieved, from 1879 to 2017.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1879
Filled flaws fail a bridge
The first Tay Rail Bridge in Scotland fell in a storm on 28 December 1879 while a passenger train was crossing it. Its cast iron columns came from a foundry at Wormit, and by Wikipedia's account of the court of inquiry, the foreman added metal afterwards to lugs that had cast badly and filled blowholes and other small flaws with a paste called "Beaumont egg" before using the castings. A later reappraisal by Peter Lewis and Ken Reynolds of the Open University traced the collapse to fatigue cracks that started at tapered holes in the cast iron lugs holding the bracing bars; cracks had been seen in inspections, but the repairs were inadequate.
LessonA flaw hidden under filler is still a flaw, and cast iron loaded in tension can break suddenly.
Credit
the court of inquiry under Henry Cadogan Rothery; Peter Lewis and Ken Reynolds
1988
Software reads wheel X-rays
H. Boerner and Helmut Strecker of the Philips research laboratory in Hamburg described several ways of finding flaws automatically in X-ray images of cast aluminum wheels, by sorting each pixel as sound or flawed from features of its neighborhood, and reported pilot tests. Their paper appeared in IEEE Transactions on Pattern Analysis and Machine Intelligence.
Credit
H. Boerner and H. Strecker, Philips GmbH, Hamburg
2000
Readers disagree on grades
Foundries grade radiographs by comparing them by eye with ASTM's graded reference radiographs (E186 and E446 for steel castings), and buyer and seller agree in advance which grade passes. Kent Carlson, Shouzhu Ou, Richard Hardin and Christoph Beckermann of the University of Iowa had five radiographers give 128 shrinkage radiographs seven ratings each. All seven agreed on both type and level for only 12.5% of them, the average 95% confidence interval was ±1.4 levels, and for 35% some ratings fell on each side of the "level 1 or better" acceptance line. A later Iowa method, in which the reader measures the indications instead of matching pictures, gave intervals of ±0.11 to ±0.38 levels on three test radiographs.
LessonThe same casting could be passed by one inspector and rejected by another.
beckermann.lab.uiowa.edubeckermann.lab.uiowa.edustore.astm.org
Credit
University of Iowa; Steel Founders' Society of America
2013
Graded pictures go digital
ASTM International published E2868, digital reference images for steel castings up to 2 in. (50.8 mm) thick, showing the categories, types and severity levels of flaws for X-rays from about 250 kV to 4 MV and for iridium-192 and cobalt-60 sources. Inspectors reading digital radiographs on a screen grade against these images; for film, ASTM still points to the E446 reference radiographs. The earliest version ASTM lists is E2868-13.
Credit
ASTM International
2017
Inspection data falsified
Kobe Steel found in 2017 that staff had falsified inspection data so that products appeared to meet customer specifications, and had made up results for products never measured. At its Daian plant, which makes aluminum cast and forged parts for aircraft and rolling stock, 67 customers were told on 26 October 2017 and 4 more later, according to findings of an Independent Investigation Committee that Kobe Steel published on 6 March 2018. Kobe Steel says it will automate testing and end data entry by one person alone.
LessonAn inspection is only as good as its record, and a record one person types in alone can be changed.
Credit
Kobe Steel's Independent Investigation Committee
Watch for
Robots carry the X-ray tube
Fraunhofer IIS showed RoboCT at the Euroguss trade fair in Nuremberg in January 2026. Two or more robots move the X-ray tube and detector around large castings, so parts too big for an inspection cabinet can be X-rayed and CT-scanned (computed tomography, which builds a 3D image from many X-ray views), and its ISAR software finds flaws automatically in both. Its Cast Control software links the results to casting process data to trace the causes of scrap. Fraunhofer gives no detection rates.
Achieved whena foundry reports routine use on production castings, with detection and false-alarm rates checked by someone outside Fraunhofer.
Claimed by
Fraunhofer IIS, January 2026
Open
01
Paired X-ray and CT set
An openly licensed set of real die-cast parts, each with X-ray images from several angles and a CT scan of the same part. The pores found in the CT scan are labeled in 3D and projected onto the X-ray images, so every flaw in the X-ray images has a checked answer. Add simulated X-ray images of the same parts and a baseline flaw detector trained only on the simulated images and scored on the real ones.
The briefCredit · unclaimed02
A machine grader under test
Software that grades digital radiographs of steel castings for shrinkage type and severity level, put through the same repeatability and reproducibility study (a gage R&R, which measures how much ratings vary between repeats and between raters) that the University of Iowa ran on human radiographers, with people and software rating the same images.
The briefCredit · unclaimed
Open 01 · the brief
Paired X-ray and CT set
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An openly licensed set of real die-cast parts, each with X-ray images from several angles and a CT scan of the same part. The pores found in the CT scan are labeled in 3D and projected onto the X-ray images, so every flaw in the X-ray images has a checked answer. Add simulated X-ray images of the same parts and a baseline flaw detector trained only on the simulated images and scored on the real ones.
Why it is still open
S. Bosse, D. Lehmhus and S. Kumar (2024) name the missing ground truth as one of three main problems in training flaw detectors for die castings, and found that simulated images give exact labels but "cannot be used directly to predict defects in manufacturing processes"; their data is available only on request. GDXray holds X-ray images without CT scans, under a research-only license. In November 2026 searches of Zenodo for "casting x-ray" and "radiograph casting" returned no records, and Hugging Face held only copies of GDXray.
Built with what exists
A commercial CT scanning service, so the team handles no radiation source; gVirtualXray, the open-source X-ray simulation library Bosse's group used; 3D Slicer to label pores in the CT volumes; PyTorch with a U-Net segmentation model; Zenodo to publish under CC BY 4.0.
First step
Ask one die caster for 20 parts its X-ray line has already rejected, with their X-ray images, and get prices for CT scans of five of them.
Done looks like
At least 100 parts from at least two die casters, each with X-ray images from at least four angles and a CT scan, published under an open license, with the simulation-trained detector's detection rate by pore size on the real images, scored against the CT labels.
Open 02 · the brief
A machine grader under test
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Software that grades digital radiographs of steel castings for shrinkage type and severity level, put through the same repeatability and reproducibility study (a gage R&R, which measures how much ratings vary between repeats and between raters) that the University of Iowa ran on human radiographers, with people and software rating the same images.
Why it is still open
The Iowa group showed in 2000 that human ratings spread by about ±1.4 levels. Its attempt that year to set the level from measured defect area worked only when the shrinkage class was already known, and its later standard still relies on a person measuring the indications. A 2025 master's thesis by Umut Can Erkan at Middle East Technical University built an X-ray data set of aluminum castings annotated to ASTM standards and compared flaw detectors on it, but its record does not say the set is public and reports no comparison with human readers. Crossref searches in November 2026 found deep-learning papers that detect casting flaws, but none that grades ASTM severity levels and reports a gage R&R against human readers.
beckermann.lab.uiowa.edubeckermann.lab.uiowa.eduopen.metu.edu.tr
Built with what exists
ASTM E2868 digital reference images, sold by ASTM, as the grading scale; digital radiographs that a steel foundry has already graded; PyTorch with a pretrained image model; the ss.rr gage R&R function in the R package SixSigma.
First step
Ask the Steel Founders' Society of America whether the radiographs and ratings from the Iowa studies can be released for research, and buy the E2868 image set.
Done looks like
100 radiographs from a foundry whose images were not used in training, each rated three times by the software (with shifts and contrast changes between runs) and three times by each of five certified radiographers, with the confidence interval in severity levels published for the software and for the people.