Industries / Industrial equipment and services / Industrial services / Commercial diving
Welding and repair
Underwater welding and repair joins, patches and taps into steel below the waterline, either wet, with the arc burning in open water, or dry, inside a gas-filled chamber clamped round the work, and fits clamps and connectors where welding is not wanted.
5
Milestones achieved, from 1932 to 2012.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1932
Electrodes that burn under water
Konstantin Khrenov developed electrodes for underwater welding in 1932 and tested them in the Black Sea. Arc welding under water was used in the mid-1930s to repair the steamship Ussuri and to raise the steamship Boris, and the German welding society DVS dates the Boris salvage, which it calls the first practical use of wet welding, to 1936. In March 1942 Khrenov continued the work in a special laboratory at the Moscow institute of railway engineers.
Credit
Konstantin Khrenov
1983
A code for underwater welds
The American Welding Society's Specification for Underwater Welding, ANSI/AWS D3.6, was first issued in 1983. It set quality classes for underwater welds, so a buyer could name the class needed and a contractor could qualify a procedure to meet it; the trade magazine Offshore wrote in 1984 that it gave "everyone in the business a common language". Class A, the top class, caps the hardness of the heat-affected zone (the steel beside the weld that the heat changes) at 325 HV10, according to TWI.
Credit
American Welding Society
1987
A shared pipeline repair kit
From 1987 Statoil managed the Pipeline Repair System pool (PRS), repair equipment paid for by a group of companies on the Norwegian shelf: plugs, clamps, mechanical connectors, handling frames and welding habitats, chambers fitted round a pipe and filled with gas so it can be welded dry at sea-floor pressure (hyperbaric welding). By 2006 PRS covered about 10,000 km of pipelines from 8 to 44 inches across in up to 600 m of water, and Statoil engineers wrote that remotely operated dry hyperbaric gas metal arc welding, with no divers, had been qualified to 1,000 m and demonstrated at 2,500 m.
Credit
Statoil and Hydro (PRS); Kjell Edvard Apeland, Jan Olav Berge and Richard Verley, Statoil (2006 paper)
1996
Water temperature matters
Wet welders held that ordinary carbon steel electrodes avoid hydrogen cracking under the weld if the steel's carbon equivalent (a measure of how readily it hardens) is about 0.40 or less, and the 1993 edition of AWS D3.6 did not list water temperature among the variables a procedure must be qualified for. T. C. West, W. E. Mitchell and R. I. Murray reported in Welding Journal in 1996 that qualification tests showed water temperature does contribute to cracking of A516 Grade 70 steel when its carbon equivalent approaches 0.40.
LessonA variable left out of a code because it seemed too small to matter can decide whether a weld cracks.
Credit
T. C. West (Welding Engineering Services), W. E. Mitchell and R. I. Murray (Naval Sea Systems Command)
2012
Robot welds a live line
On Statoil's Åsgard field in the Norwegian Sea a remotely controlled welding robot welded a T-piece onto the Åsgard B production flowline at 265 m while gas flowed through it, and a remote drilling machine then cut into the line, an operation called a hot tap; Statoil called it a world record. Its earlier remote hot taps, on Tampen Link and in 2009 at 860 m on Ormen Lange, had used T-pieces fitted in advance. The new tie-in point was for Åsgard's subsea gas compression.
Credit
Statoil; Kjell Edvard Apeland (project manager)
Watch for
Welding robot in a harbor
DFKI, the German Research Center for Artificial Intelligence, says its UWS underwater welding robot, an underwater arm with stereo cameras, AI that finds the joint and plans the path, and flux-cored wire welding, made test welds in the Kaiserhafen III basin in Bremerhaven, in waves and murky water, of a quality comparable to its laboratory welds. The robot was on a crane-mounted frame, and the project's photographs show divers supervising the trials.
Achieved whenwelds made outside the laboratory are tested by an independent body against an AWS D3.6 class, or the robot repairs a working structure.
Claimed by
DFKI with TH Köln, Fraunhofer IGD and AMT GmbH, July 2026
Open
01
Shared table of wet welds
An open table of published wet weld test results, one row per weld: depth or pressure, water temperature and salinity, electrode and coating, base steel and its carbon equivalent, heat input, porosity, heat-affected zone hardness, bend and tensile results, and the AWS D3.6 class met. Then a model that predicts hardness and pass or fail for welds it has not seen.
The briefCredit · unclaimed02
Seams found in murky water
A public tank test of how well cameras and laser-stripe scanners locate and measure a weld groove under water as the water clouds, with no welding done: steel plates with machined grooves and joints of known shape, scanned at set turbidity levels and scored against the machined dimensions.
The briefCredit · unclaimed
Open 01 · the brief
Shared table of wet welds
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open table of published wet weld test results, one row per weld: depth or pressure, water temperature and salinity, electrode and coating, base steel and its carbon equivalent, heat input, porosity, heat-affected zone hardness, bend and tensile results, and the AWS D3.6 class met. Then a model that predicts hardness and pass or fail for welds it has not seen.
Why it is still open
The models found each use one laboratory's own welds: Omajene and colleagues at Lappeenranta University of Technology trained a neural network on bead shape (2014), and Chen and Feng predicted seam shape from camera and arc signals (2014). West and colleagues showed in 1996 that water temperature, which the code then left out, changes cracking. A search in November 2026 found no pooled table of wet weld results across laboratories.
Built with what exists
Crossref and OSTI to find papers; open-access PDFs read with PyMuPDF; a low-cost language model to pull each table into rows and quote the table it used (see the Note "The jobs agents repeat"); two people to check a sample; scikit-learn for the model.
First step
Collect 20 open-access wet welding papers and copy their result tables by hand to fix the column list.
Done looks like
At least 1,000 welds from at least 40 papers and 10 laboratories, each row traced to a page and table; two checkers agree on 95% of 100 sampled rows; a model trained on papers published before 2015 that predicts the hardness of later welds within 25 HV10 for at least 80% of them.
Open 02 · the brief
Seams found in murky water
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public tank test of how well cameras and laser-stripe scanners locate and measure a weld groove under water as the water clouds, with no welding done: steel plates with machined grooves and joints of known shape, scanned at set turbidity levels and scored against the machined dimensions.
Why it is still open
Remote and robotic welding depends on finding the joint. Yoji Ogawa showed in 1999 that a slit laser still picked out groove shape in turbid water, and T. Ekkel and colleagues at Leibniz University Hannover built a stereo camera and line laser system to measure weld seams under water (2015), but neither published scans at measured turbidity for others to test on. DFKI's 2026 harbor trials of AI joint finding gathered data in murky water that its partners are analyzing, and no public set was found.
fujipress.jpisprs-archives.copernicus.orgrobotik.dfki-bremen.de
Built with what exists
A water tank; a stereo camera and a low-power green line laser in waterproof housings on a gantry or a BlueROV2; kaolin clay to cloud the water and a turbidity meter reading in NTU (nephelometric turbidity units); OpenCV for laser-stripe and stereo baselines; a general image segmentation model as another method.
First step
Have a shop machine five V-groove plates, then scan them in clear water and compare the readings with calipers.
Done looks like
30 joints scanned at four turbidities (0, 5, 15 and 30 NTU), with position and width errors in millimetres published for each method and turbidity, all scans and the machining drawings released, and at least one method shown to hold position within 0.5 mm at 15 NTU.