BRENNEN

Industries / Industrial equipment and services / Industrial services / Non-destructive testing

X-ray and gamma ray

Radiography sends X-rays from a tube, or gamma rays from a sealed radioactive capsule, through a weld or casting onto film or a digital detector, and a hidden hole or crack shows as a darker mark because less metal stood in the way.

5

Milestones achieved, from 1922 to 2015.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1922

    X-rays enter the foundry

    Watertown Arsenal, the US Army's metalworks near Boston, began in 1922 to use radiography to find and remove defects in steel castings and to control the quality of welds, with a 280,000-volt X-ray machine for testing materials. Horace Hardy Lester, at the arsenal from 1922 to 1953, led the work; the Historic American Engineering Record credits him with improved steel castings and welded structures, and the arsenal's materials research reactor was named after him in 1960.

    loc.govtile.loc.gov

    Credit

    Horace Hardy Lester, Watertown Arsenal

  2. 1920s

    Radium checks cruiser castings

    At the US Naval Research Laboratory (NRL), Robert F. Mehl worked out how to make pictures of flaws in steel with gamma rays from radium, a naturally radioactive metal. G. E. Doan of Lehigh University wrote in 1931 that X-ray equipment of the day could practically get through only about 3.5 inches (89 mm) of steel, and that the idea of using radium was Mehl's. NRL's own history says the method was first used on suspected flaws in the sternpost castings of the US Navy's new 9,000-tonne heavy cruisers; the castings were found faulty and all ten cruisers had to be repaired. From 1946, gamma sources made in nuclear reactors, cobalt and iridium, which were stronger and cheaper than radium, replaced it, according to Iowa State University's NDE-Ed resource.

    tms.orgstore.astm.orgnde-ed.org

    Credit

    Robert F. Mehl, US Naval Research Laboratory

  3. 1984

    A lost source kills eight

    A radiography source is a small sealed capsule of radioactive metal that is wound out of a shielded container on a cable to make an exposure and wound back afterward. In 1984 a 1,100 GBq (30 curie) iridium-192 source came off its drive cable, was not returned to its container and fell to the ground at the work site. A passer-by picked up the tiny metal cylinder and took it home. Between March and June eight people, he and members of his family and relatives, died; doctors first assumed poisoning. The International Atomic Energy Agency (IAEA) reviewed the case in its 1998 report on radiography accidents, citing a US Nuclear Regulatory Commission notice that names Morocco. It found that no radiation survey was made to confirm the source was back in its container, and that the capsule itself carried no warning mark.

    LessonThe IAEA found that radiography accidents happen "primarily, because of failure to adhere to procedures", and a survey with a radiation meter after each exposure would have shown the source was loose.

    www-pub.iaea.org

    Credit

    a failure; reviewed by the IAEA in Safety Reports Series No. 7

  4. 2013

    Digital radiographs for welds

    For more than 40 years, contracts for radiography were written around the picture quality of film. The first standards for computed radiography, which uses reusable imaging plates read by a laser scanner, were published by CEN, the European standards body, and ASTM International in 2005; ASTM published its first standards for using digital detector arrays, flat panels that give an image at once, in 2010. ISO 17636-2, published in 2013, was the first ISO practice for radiographing welds with these detectors, and it set minimum image quality to match film. In 2014 Uwe Ewert and Uwe Zscherpel of BAM, Germany's Federal Institute for Materials Research and Testing, wrote that fast serial inspection of small castings was "almost completely dominated" by digital radiography, while auditors and managers at large plants and construction sites still asked for film.

    LessonDigital radiography spread first where short exposures paid off, in serial casting checks, and slowest where contracts and audits were built on film.

    ndt.netstore.astm.org

    Credit

    ISO, CEN and ASTM International; account by Uwe Ewert and Uwe Zscherpel of BAM

  5. 2015

    A shared X-ray image library

    Domingo Mery and seven co-authors published GDXray, a database of X-ray images for nondestructive testing that includes castings, such as a series of 72 images of an aluminum wheel, and welds. The weld images came from Uwe Zscherpel at BAM. Its site now lists more than 21,100 images in the larger GDXray+. It gave researchers outside factories a common set on which to test automatic defect recognition, software that marks flaws in X-ray images, and by November 2026 Crossref counted 457 publications citing the paper. The images are free for research and teaching only; redistribution and commercial use are prohibited.

    domingomery.ing.uc.clapi.crossref.org

    Credit

    Domingo Mery and colleagues; BAM

Watch for

  1. CT flags bad battery cells

    Computed tomography (CT) builds a 3D picture of the inside of a part from X-ray images taken at many angles. Lumafield, a US maker of CT scanners, says it scanned 1,054 cylindrical 18650 lithium-ion cells from ten brands, from established makers to rewrapped and low-cost marketplace labels, and found negative anode overhang, an electrode misalignment that raises the risk of an internal short, in 33 of them, all from low-cost or counterfeit brands. It says its Triton factory scanner takes under 5 seconds a cell, more than 720 cells an hour.

    Achieved whenan outside laboratory scans or cuts open cells of the same brands and reports a similar defect rate.

    lumafield.com

    Claimed by

    Lumafield, September 2025

  2. A model trained on CT

    Lumafield says its Quality Agent runs on a foundation model, a large AI model trained on "the largest library of industrial CT data anywhere", that learns what a good part looks like and flags anything that departs from it, instead of looking for a list of named defects. A review of the launch on the Pebblous blog notes that the announcement gave no detection rate, false-alarm rate or named customer; the evidence offered was that the model sorted the same 1,054 battery cells by maker.

    Achieved whena customer or an outside laboratory publishes the share of real defects it caught, and its false alarms, on parts the company did not choose.

    lumafield.comblog.pebblous.ai

    Claimed by

    Lumafield, September 2026

Open

  1. 01

    Technique sheets checked by software

    A program that reads a radiographic technique sheet, the written plan for one exposure, and checks it against the rules of ASME Section V Article 2 or ISO 17636-1 and 17636-2, quoting the clause behind each finding. A sheet gives the source or tube and its focal spot size, the distances, the film or detector class, the image quality indicator (a wire or plate of known size laid on the part to prove the picture is sharp enough) and the number of shots around a pipe. This is paper work only: no radiation source or X-ray tube is used.

    The brief
    Credit · unclaimed
  2. 02

    Radiography accidents since 1998

    A public table of every published industrial radiography accident or overexposure since the IAEA's 1998 review, taken from regulators' event reports in at least five countries. Two people code each event separately to the review's eight primary causes: inadequate regulatory control, failure to follow procedures, inadequate training, inadequate maintenance, human error, equipment malfunction, design flaws and willful violation. Then compare the mix of causes with 1998. The work is reading only: no one handles a source.

    The brief
    Credit · unclaimed

Open 01 · the brief

Technique sheets checked by software

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A program that reads a radiographic technique sheet, the written plan for one exposure, and checks it against the rules of ASME Section V Article 2 or ISO 17636-1 and 17636-2, quoting the clause behind each finding. A sheet gives the source or tube and its focal spot size, the distances, the film or detector class, the image quality indicator (a wire or plate of known size laid on the part to prove the picture is sharp enough) and the number of shots around a pipe. This is paper work only: no radiation source or X-ray tube is used.

Why it is still open

Which rules apply depends on the contract, and Ewert and Zscherpel of BAM found in 2014 that the ISO, CEN, ASTM and ASME practices for digital radiography set different requirements for image quality and equipment. Study guides teach the checks by hand, such as ASME Section V's limit on geometric unsharpness, the blur from the size of the source, of 0.020 inch (0.51 mm) for material under 2 inches (51 mm) thick. A November 2026 search found no tool that checks a whole technique sheet against these rules.

ndt.netopen-exam-prep.com

Built with what exists

Python; licensed copies of ASME Section V and ISO 17636; each rule written as a small function; a language model that reads scanned forms, such as Anthropic's Claude, calling those functions and citing the clause for each finding, the way Cursor's example code reviewer cites the note behind each finding (see the Note "Findings that cite notes").

  1. First step

    Write 20 technique sheets from the worked examples in a radiography textbook, plant errors in 10 of them, and code the geometric unsharpness and image quality indicator checks first.

  2. Done looks like

    200 technique sheets written by Level III radiographers, 100 of them with one planted error each, released with the answer key; the program catches at least 95 of the 100 errors and wrongly flags no more than 5 of the 100 correct sheets, with the scores of two Level III reviewers on the same sheets published beside it.

Open 02 · the brief

Radiography accidents since 1998

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public table of every published industrial radiography accident or overexposure since the IAEA's 1998 review, taken from regulators' event reports in at least five countries. Two people code each event separately to the review's eight primary causes: inadequate regulatory control, failure to follow procedures, inadequate training, inadequate maintenance, human error, equipment malfunction, design flaws and willful violation. Then compare the mix of causes with 1998. The work is reading only: no one handles a source.

Why it is still open

The IAEA review described a selection of accidents grouped by cause but gave no full count, and noted that accidents are mostly known to the few countries with the regulatory infrastructure to collect them. A 2015 systematic review by Kevin Coeytaux and colleagues counted 634 reported radiation overexposure accidents worldwide from 1980 to 2013, from every use of radiation, and described them by region and type, not by the IAEA's causes. A November 2026 search found reviews of US events from the late 1980s and 1990s and US yearly counts, but no worldwide table of radiography events coded to the 1998 causes.

www-pub.iaea.orgebi.ac.uk

Built with what exists

Public event reports from regulators such as the US Nuclear Regulatory Commission, Australia's ARPANSA, Canada's CNSC and the UK's HSE, and IAEA publications; a language model that drafts a cause code for each report together with the passage behind it, checked by a person, as in Enveda's Claude-based record reviewer (see the Note "Seven more flagged", a draft); Python and pandas; Cohen's kappa, a standard measure of agreement between two coders.

  1. First step

    Have two people code 30 US industrial radiography event reports from one year separately, and measure how often they agree.

  2. Done looks like

    At least 300 events coded by two coders with a kappa above 0.7, released with a link to each report, and a table comparing each cause's share with the 1998 review.