BRENNEN

Industries / Mining and materials / Steel and metals / Forging

Open-die forging

Open-die forging squeezes hot metal between flat or simply shaped tools that do not enclose it, turning and feeding the piece between strokes to draw out shafts, rings and blocks, often from ingots of hundreds of tonnes.

5

Milestones achieved, from 1841 to 2021.

1

To watch for: announced, not yet confirmed outside.

3

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1841

    A hammer lifted by steam

    A steam hammer raises a heavy block with a steam cylinder and lets it fall on the metal, so the size and pace of the blows no longer depend on a water wheel. At Le Creusot in France, François Bourdon built one for the Schneider brothers' ironworks, where it forged shafts for marine steam engines; ASME, the American Society of Mechanical Engineers, credits Bourdon and Eugène Schneider with "the construction, patenting, and practical employment of the first true steam hammer in 1841." James Nasmyth, whose drawing of a steam hammer is dated 24 November 1839, wrote that Schneider and Bourdon had copied it from his Scheme Book on a visit to his works near Manchester, that he found their hammer at work in April 1842, and that he patented his own in June 1842. ASME calls the invention one of the engineering "firsts" most "clouded in controversy."

    asme.orgarchive.org

    Credit

    François Bourdon and Eugène Schneider, Schneider & Cie; James Nasmyth, Bridgewater Foundry, Patricroft

  2. 1862

    Pressing instead of hammering

    A hydraulic forging press squeezes the metal slowly with water pressure instead of striking it. John Haswell patented one in 1861 and showed it at the London Exhibition of 1862; his obituary in The Engineer (1897) called it "the first of the kind in England" and said small articles had been stamped by hydraulic pressure before. Joseph Whitworth replaced the steam hammer with the hydraulic press for forging gun steel, and the US Gun Foundry Board, after visiting his Openshaw works in Manchester in 1883, judged that his system "surpassed all other methods of forging."

    gracesguide.co.ukgracesguide.co.uk

    Credit

    John Haswell; Joseph Whitworth

  3. 1902

    The biggest hammer scrapped

    Hammers grew with ingots: Krupp's 50-ton "Fritz" at Essen in 1861, Schneider's hammer of 80 to 100 tons at Le Creusot in 1877, and a 125-ton hammer that the Bethlehem Iron Company built in 1891 under patent rights bought from Schneider (the tons are ASME's ratings of each hammer's blow). Bethlehem demolished its hammer in 1902, and the Creusot hammer worked until 1930; both lost out to hydraulic and mechanical presses. ASME gives the reason: a press applies its force slowly and as strongly at the end of the stroke as at the start, while hammering "often altered only the outer surfaces, leaving large forgings with damaging internal stresses."

    LessonFor large forgings, how deep the force reaches into the metal matters more than how hard the blow lands.

    asme.org

    Credit

    Schneider & Cie; Bethlehem Iron Company; Krupp

  4. 2008

    One forge, most reactors

    A reactor pressure vessel is assembled from forged rings and heads, each cut from a steel ingot that can weigh hundreds of tonnes. World Nuclear News reported in December 2008 that around 130 reactor pressure vessels from Japan Steel Works (JSW) were in service, that JSW claimed 80% of the world market for large nuclear forgings, and that its largest press at Muroran, of 14,000 tonnes, took 600-tonne ingots; JSW said it would triple capacity from four reactor vessels a year to twelve by mid-2012. The World Nuclear Association reports Westinghouse's view that the largest AP1000 parts need at least a 15,000-tonne press taking 350-tonne ingots, and says nothing in North America approaches the heavy forges of Japan, China, South Korea, France and Russia.

    world-nuclear-news.orgworld-nuclear.org

    Credit

    Japan Steel Works

  5. 2021

    A government buys its forge

    On 28 July 2021 the UK Ministry of Defence offered to buy Sheffield Forgemasters for £2.56 million plus its debt, calling it "the only available manufacturer" able to make large, high-integrity castings and forgings from specialist steels for its programs. The purchase was completed on 19 August 2021, and the ministry said it would invest up to £400 million over ten years in plant, equipment and buildings.

    gov.uk

    Credit

    UK Ministry of Defence; Sheffield Forgemasters

Watch for

  1. The UK's largest open-die press

    Sheffield Forgemasters says it will install a 13,000-tonne open-die press built by Mitsubishi Nagasaki Machinery, served by a 450-tonne rail-bound manipulator (a machine that grips the ingot and turns and feeds it between strokes) and a second manipulator that changes tools automatically. The press belongs to a £1.3 billion rebuilding program for the Ministry of Defence-owned company; installation is due to start in July 2027 and take 12 months.

    Achieved whenthe press is commissioned and forging production ingots, as confirmed by a customer, the Ministry of Defence or the trade press.

    sheffieldforgemasters.com

    Claimed by

    Sheffield Forgemasters, September 2025

Open

  1. 01

    Stroke logs from a camera

    A program that watches ordinary video of an open-die press at work and writes a log of every stroke (when it fell and where along the workpiece it landed) and of every turn or feed by the manipulator, released with a set of hand-annotated videos to test it on.

    The brief
    Credit · unclaimed
  2. 02

    A register of heavy presses

    An open table of the world's open-die presses of 80 MN (about 8,000 tonnes) and up and closed-die presses of 300 MN and up, giving owner, site, rated force in the owner's units and in meganewtons (MN), largest ingot, year of commissioning and status, with every figure linked to a dated source.

    The brief
    Credit · unclaimed
  3. 03

    One test for pass planners

    A public set of pass-schedule tasks for open-die forging (starting ingot, target bar, press force limit, largest allowed bite) with a scoring script that runs on an open simulator, plus a check of the best schedules on cold plasticine bars squeezed between flat dies in a university testing machine.

    The brief
    Credit · unclaimed

Open 01 · the brief

Stroke logs from a camera

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 watches ordinary video of an open-die press at work and writes a log of every stroke (when it fell and where along the workpiece it landed) and of every turn or feed by the manipulator, released with a set of hand-annotated videos to test it on.

Why it is still open

Niklas Reinisch and colleagues write that an open-die forging can take "several hundred strokes" and that pass schedules, the planned sequence of strokes and turns, are "often only based on experience or simple mathematical models." Models that plan or check schedules need stroke-by-stroke records, and the open JAX-FORGE simulator replays runs of 12 and 17 hits recorded on instrumented equipment. Research cells gather such records with their own sensors: Roy Rechenberg and colleagues at TU Bergakademie Freiberg (2025) describe a robot forging cell that measures the workpiece by 3D laser scanning and thermal imaging, and a 2020 RWTH Aachen paper is titled "Action Discretization for Robot Arm Teleoperation in Open-Die Forging". Searches of Crossref, GitHub and Hugging Face in November 2026 found no published method or dataset for logging strokes from ordinary video of a working forge, and the two Hugging Face datasets named for forging hold almost nothing.

doi.orggithub.comcmms.agh.edu.pl

Built with what exists

Public videos of open-die forging posted by forges and press makers, used with their permission; Meta's SAM 2 model to follow the glowing workpiece from frame to frame; a vision-language model reading sampled frames to label strokes and turns; Label Studio for the hand annotations.

  1. First step

    Collect 30 public videos of open-die forging, annotate one by hand with stroke times, and write to two forges asking to use their footage.

  2. Done looks like

    20 videos from at least five forges, annotated stroke by stroke by two people; on videos it was not tuned on, the program finds at least 95% of strokes and 90% of manipulator turns, each timed within half a second, with every miss published.

Open 02 · the brief

A register of heavy presses

Credit · unclaimed

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

What to build

An open table of the world's open-die presses of 80 MN (about 8,000 tonnes) and up and closed-die presses of 300 MN and up, giving owner, site, rated force in the owner's units and in meganewtons (MN), largest ingot, year of commissioning and status, with every figure linked to a dated source.

Why it is still open

The figures are scattered and their units clash: the press at Weber Metals in California is 540 MN and 60,000 short tons to its builder, SMS group, and 54,000 tons to its owner, Otto Fuchs. The World Nuclear Association's table of heavy forges covers nuclear work only and gives press sizes without dates and with question marks. A 2026 working paper by Matthew Player, Michael Tadrous, Matthew Zychowicz and Goran Calic models 16 facilities in 9 countries that make reactor vessel and steam generator forgings, but only that class of forge. Searches of GitHub, Hugging Face, Crossref and Wikipedia in November 2026 found no open, row-by-row register.

world-nuclear.orgdoi.org

Built with what exists

Company histories and annual reports, government notices and trade press; an AI model to pull candidate rows and citations out of those documents, each row checked by a person, in the way Enveda's Claude-based reviewer flags issues in long records (see the Note "Seven more flagged"); a spreadsheet published on GitHub.

  1. First step

    Turn the World Nuclear Association table and the presses named in this file into the first 30 rows.

  2. Done looks like

    At least 60 presses in at least 12 countries, each force converted to MN with the conversion shown, each row carrying a source and access date, a second person's check of 20 rows picked at random, all released under a CC BY 4.0 license.

Open 03 · the brief

One test for pass planners

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 pass-schedule tasks for open-die forging (starting ingot, target bar, press force limit, largest allowed bite) with a scoring script that runs on an open simulator, plus a check of the best schedules on cold plasticine bars squeezed between flat dies in a university testing machine.

Why it is still open

Planners exist, but each has been tested by its own authors on their own tasks. Reinisch and colleagues trained a double deep Q-learning agent, a kind of reinforcement learning, on one industrial-scale process and checked it with one forging experiment in 2021; Christian Linder, Martin Feistle and Syed Irtiza Ali presented a method combining reinforcement learning and simulated annealing in April 2026. JAX-FORGE now offers an open simulator, and the GitHub repository forging-control trains neural network controllers for a forging press. Crossref and GitHub searches in November 2026 found no shared task set with public scores. Łukasz Wójcik, Konrad Lis and Zbigniew Pater measured plasticine's flow curves at 0 to 20 °C for use in physical models of metal forming, so it can stand in for hot steel in cold tests.

doi.orgdoi.orgdoi.org

Built with what exists

JAX-FORGE, a GPU finite element simulator built on JAX-FEM; Python with Stable-Baselines3 for reinforcement learning; an AI coding agent that writes schedules as a third entrant; plasticine; a universal testing machine with flat dies, run by laboratory staff.

  1. First step

    Write five tasks from published ingot and bar sizes and run JAX-FORGE's demonstration case on the first.

  2. Done looks like

    20 tasks and three planners (a textbook rule, a reinforcement learning agent and an AI agent) scored on passes used, force-limit breaches and final shape error; the five best schedules repeated three times each in plasticine, with final dimensions within 2 mm of the simulation, all published.