BRENNEN

Industries / Mining and materials / Steel and metals / Forging

Inspection

Inspection checks that a finished forging is sound inside and on its surface, the right size and backed by true records, before it goes into a turbine, a reactor or an engine.

5

Milestones achieved, from 1956 to 2020.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1956

    Flaws inside a burst shaft

    The low-pressure spindle, the large forged shaft, of a 165,000 kW steam turbine at Ridgeland burst during a routine overspeed-trip test. In a 1956 paper H. D. Emmert of Allis-Chalmers reported that the cause was flakes (fine internal cracks) or thermal cracks formed in the shaft during heat treatment, which gave the nickel-molybdenum-vanadium steel, prone to brittle failure at notches in heavy sections, the notches it needed to burst. The case, he wrote, focused attention on the metallurgy, production and inspection of large turbine-spindle forgings.

    doi.org

    Credit

    H. D. Emmert, Steam Turbine Department, Allis-Chalmers Manufacturing Company

  2. 1974

    A rotor bursts at Gallatin

    On 19 June 1974 a turbine rotor at the Tennessee Valley Authority's Gallatin No. 2 unit burst into dozens of pieces at about 3,400 rpm during a cold start, after 106,000 hours in service. It had been forged in 1954 from an air-melted ingot and held a cluster of manganese sulfide inclusions near its bore, the hole along its axis. Ultrasonic inspection from inside the bore before service was then required only for rotors heat-treated at 1,100 °C, and this one had been treated at 955 °C, so the cluster went unseen while a crack grew from it. US power companies then inspected the bores of all their rotors, and the Electric Power Research Institute (EPRI) wrote the SAFER code for deciding whether to run or retire a rotor.

    Lessona rule that tied inspection to one heat-treatment route left other forgings unchecked.

    shippai.org

    Credit

    Tennessee Valley Authority; EPRI; case compiled by Japan's Failure Knowledge Database

  3. 2016

    Forge records questioned

    After France's nuclear safety authority, ASN, judged an October 2015 review insufficient, Areva re-examined the records of its Le Creusot forge. On 26 April 2016 it reported "irregularities" in manufacturing checks on about 400 parts made there since 1965, about 50 of them thought to be in service in French nuclear plants; ASN described them as inconsistencies, modifications or omissions in the production files. Areva said nothing found so far put the parts' mechanical integrity in doubt, and ASN required every anomaly to be assessed.

    Lessona forging's paper trail is part of what has to be inspected.

    world-nuclear-news.org

    Credit

    ASN; Areva

  4. 2017

    Carbon left in vessel heads

    A steel ingot has zones that naturally hold excess carbon, and forgers cut them away. ASN said in June 2017 that Areva's Creusot Forge plant had not removed enough of such a zone when forging the lower and closure heads of the Flamanville 3 reactor pressure vessel, so steel with excess carbon, and lower toughness (resistance to crack growth), remained in the middle of both. Tests on parts made the same way, run mostly by laboratories independent of Areva, found toughness below the design value. ASN accepted the vessel with extra inspections, but ruled that the closure head could not stay in service beyond the end of 2024 because inspecting it in service might not be feasible. A year earlier ASN had said that steam generator channel heads in 18 reactors, forged by Creusot Forge and by JCFC in Japan, could hold a similar zone.

    french-nuclear-safety.frregulation-oversight.asnr.fr

    Credit

    ASN; Areva NP; EDF

  5. 2020

    A crack traced to macrozones

    On 30 September 2017 the fan hub, the forged titanium disk that holds the fan blades, broke on an Engine Alliance GP7270 engine of an Air France Airbus A380 over Greenland, and the crew diverted to Goose Bay, Canada. After airborne and ground radar searches, a hub fragment was dug out of the ice sheet in June 2019. The BEA, France's air accident investigator, reported in September 2020 that a fatigue crack had started 1.4 mm below the surface of a blade slot, in a macrozone, a patch of similarly oriented crystals that forms when the ingot is converted to billet and that forging only reduces. Titanium alloys creep at room temperature under the load held through a flight, which speeds fatigue (cold dwell fatigue); the hub had flown 3,534 cycles, and the BEA listed the lack of any non-destructive method to detect unusual macrozones among the causes.

    Lessona forging can meet its specification and still hold a weak zone that current inspection cannot see.

    bea.aero

    Credit

    BEA, with the Geological Survey of Denmark and Greenland (GEUS), ONERA and Pratt & Whitney

Watch for

  1. Cameras grade forgings

    SMS group and Hirschvogel, a forging company, said in March 2026 that they are replacing manual final-inspection stations for closed-die forgings at Hirschvogel's Denklingen site in Germany with automatic testing for surface defects and shape errors, built with the image processing firm Sightwise. High-resolution cameras photograph each part, and the defects found are sorted against a digital catalog; the analysis models are trained on photo-like images generated from CAD models, which takes 6 to 8 hours. No detection rates or start date were given.

    Achieved whenHirschvogel or an outside auditor publishes missed-defect and false-reject rates against manual inspection over months of production.

    aist.org

    Claimed by

    SMS group and Hirschvogel, March 2026

  2. A new Flamanville head

    EDF says the replacement closure head for Flamanville 3, made by Framatome at its Saint-Marcel plant and weighing about 99 tonnes before equipping, arrived on site on 16 June 2026 to be fitted during the reactor's first full inspection outage. The reactor was disconnected from the grid on 26 September 2026 for that outage; the 89 control-rod drive mechanisms are to be moved to the new head and tested, and the old head stored on site for two years.

    Achieved whenthe regulator, now named ASNR, allows the reactor to restart with the new head in place.

    edf.fredf.fr

    Claimed by

    EDF, June 2026

Open

  1. 01

    Forging photos across months

    An openly licensed set of photos of finished closed-die forgings, taken at the final-inspection stations of at least three forges over at least six months, with surface defects marked by class (laps and folds, cracks, underfill, mismatch between the die halves, scale pits) and each rejected part confirmed by the forge's own inspectors. Add a test that trains a model on early months at one forge and scores it on later months and other forges.

    The brief
    Credit · unclaimed
  2. 02

    Etched sections rated by machine

    A public set of photos of macroetched cross-sections of forgings (slices ground flat and etched with acid so that segregation, laps, cracks and porosity show to the eye at low magnification), each rated by two metallurgists, and a model that rates a new photo for the same conditions.

    The brief
    Credit · unclaimed

Open 01 · the brief

Forging photos across months

Credit · unclaimed

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 photos of finished closed-die forgings, taken at the final-inspection stations of at least three forges over at least six months, with surface defects marked by class (laps and folds, cracks, underfill, mismatch between the die halves, scale pits) and each rejected part confirmed by the forge's own inspectors. Add a test that trains a model on early months at one forge and scores it on later months and other forges.

Why it is still open

Tatsuya Yamazaki and Akito Fukui of Niigata University wrote in 2020 that forged parts are checked by visual inspection and that image processing and laser measurement had hardly been compared. A graduation project on GitHub, using images supplied by a Japanese forging maker, reports 100% accuracy on the month its model trained on and about 50% on the three months after, with nearly every defective part passed as good. SMS group trains its models on synthetic images and has published no accuracy. Searches of Hugging Face, Zenodo and GitHub in November 2026 found no public set of forging-defect photos.

github.comdoi.org

Built with what exists

A fixed camera with a ring light at each station, or a phone on a stand; Label Studio for marking defects; open detection models such as Ultralytics YOLO; a vision-language model as a second reader; Hugging Face for hosting. Parts are photographed cold, after inspection.

  1. First step

    Ask one forge for 200 photos of parts its inspectors have already sorted as good or rejected, from two months at least three months apart, and rerun the month-to-month test.

  2. Done looks like

    At least 5,000 photos from at least three forges over at least six months, at least five defect classes with written definitions, and a baseline whose missed-defect rate at a fixed false-reject rate is published for a held-out later month and a held-out forge.

Open 02 · the brief

Etched sections rated by machine

Credit · unclaimed

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

What to build

A public set of photos of macroetched cross-sections of forgings (slices ground flat and etched with acid so that segregation, laps, cracks and porosity show to the eye at low magnification), each rated by two metallurgists, and a model that rates a new photo for the same conditions.

Why it is still open

ASTM E381, the method for macroetch testing of steel bars, billets, blooms and forgings, rates specimens against a graded series of photographs, by eye; the conditions it covers include segregation, laps, seams, cracks, porosity, bursts, pipe and flakes. Crossref and GitHub searches in November 2026 for software that rates macroetched sections of forgings returned only the standards themselves. Nearby work is on cast steel: a team at the Federal University of Ouro Preto, Brazil, classed central segregation in etched slab sections with a neural network trained on 150 images and tested on 30, and Xiaolu Wei and colleagues (April 2026) predicted carbon content from low-magnification sections of 600 mm 42CrMo billets, with their data available only on request. ASTM's site says that using AI on its standards is prohibited, so a model cannot be trained on its reference photographs, and new photos have to be collected.

store.astm.orgsol.sbc.org.broaepublish.com

Built with what exists

Photos of sections that university and forge laboratories have already cut and etched, so the acid work stays in those labs; a flatbed scanner or a camera with a scale bar; Label Studio; a small image classifier and a vision-language model as baselines; Cohen's kappa, a measure of agreement beyond chance, to compare raters.

  1. First step

    Ask two university metallurgy labs to let you photograph their archived etched forging sections, and draft the labs' own rating sheet for the conditions metallurgists look for, such as segregation, laps, cracks, porosity and flakes.

  2. Done looks like

    At least 600 etched sections from at least four laboratories, each rated by two metallurgists, published with the agreement between the two raters and between the model and each rater.