Industries / Industrial equipment and services / Industrial services / Commercial diving
Training and medicine
This domain covers how commercial divers are trained and certified, and the medicine of the work: how to decompress safely, how to check divers' health, and what the work does to the body over a career.
7
Milestones achieved, from 1839 to 2026.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1839
Soldiers learn from divers
Colonel Charles Pasley of the Royal Engineers, who had taken over the clearing of the wreck of HMS Royal George from the anchorage at Spithead, hired the Whitstable divers George Hall and Hiram London in 1839 and had them train his own Sappers and Miners. The civilians shared the work in 1840 and left in the third year, and by 1842, John Bevan writes, Pasley "had established a fully operational diving capability under his command." Three years earlier John Deane had published "Method of Using Deane's Patent Diving Apparatus", which Bevan calls the first ever diving manual.
LessonOn an earlier job, the wreck of the brig William in the Thames, Pasley had been advised he could use untrained Sappers and Miners as divers, and one died tangled in the wreck.
Credit
Charles William Pasley; George Hall and Hiram London
1908
Ascent in stages
Working at the Lister Institute for the Admiralty's Committee on Deep Diving, A. E. Boycott, G. C. C. Damant, a Royal Navy inspector of diving, and J. S. Haldane compressed goats in a steel chamber and worked out how fast the body takes up and gives off nitrogen. They showed that a diver could rise quickly until the absolute pressure was halved and must rise slowly after that, with stops set so that the excess nitrogen in any part of the body never exceeded double the nitrogen pressure of the air breathed. Their tables were already those "in use for divers in the Royal Navy".
Credit
A. E. Boycott, G. C. C. Damant and J. S. Haldane
1976
Bubbles heard before bends
M. P. Spencer used ultrasonic Doppler, a probe that turns blood flow into sound, to listen for venous gas emboli (bubbles in the veins) after chamber exposures from 233 ft for 7 minutes to 25 ft for 720 minutes, and after open-sea dives to 165 ft for 10 minutes. No case of the bends developed without bubbles first being heard over the heart, and some divers formed bubbles far more readily than others. He also found that the common belief that any length of time at 30 ft was safe did not hold.
LessonBubble counts gave a measure that can be taken after every test dive, long before enough people are hurt to count cases of the bends.
Credit
M. P. Spencer
1982
Schools agree standards
The International Diving Schools Association (IDSA) was formed in 1982 to develop common international diving standards. It publishes International Diver Training Standards for offshore, inshore and inland divers and for supervisors and technicians, and keeps a Table of Equivalence that maps the national training schemes some governments set against one another.
Credit
International Diving Schools Association
1984
Risk as a probability
Decompression tables had been built on a critical value: a profile was either safe or unsafe. P. K. Weathersby, L. D. Homer and E. T. Flynn applied maximum likelihood, a standard statistical method, to fit models to the outcomes of real decompressions in rats and men, which gave a consistent way to estimate the chance of decompression sickness for a profile and the uncertainty in that estimate.
Credit
P. K. Weathersby, L. D. Homer and E. T. Flynn
2017
A training centre closes
On 14 September 2017 The Underwater Centre Tasmania went into liquidation. It was the only centre in the southern hemisphere accredited under the Australian Diver Accreditation Scheme (ADAS) and recognised by the International Marine Contractors Association (IMCA) to train divers at its highest level, and Australia's only civilian centre at the level below. David Smart, a diving physician, wrote that Australia's 2012 model safety laws treated diving other than construction diving as low risk, opening the way to shorter industry courses not recognised elsewhere, and that in Tasmania's fish farms in the 1980s, before training to the ADAS benchmark, divers were injured or fell ill at over 50 times the current rate.
LessonIn Smart's view, training standards erode when the law stops treating diving as one high-risk trade.
Credit
David Smart, Royal Hobart Hospital, for the account
2026
Ten thousand bubble recordings
In March 2026 a group funded by the US Office of Naval Research released 10,099 Doppler recordings of venous gas emboli from up to 311 military and scientific divers, digitised from old tapes over four years and paired with grades and dive data. They say it is the first public collection of the original audio, not only the grades investigators gave it; it includes only 12 cases of decompression sickness.
Credit
S. Lesley Blogg, Arian Azarang, Peter Lindholm and colleagues, with the University of California San Diego, the University of North Carolina at Chapel Hill, Divers Alert Network, Duke University, the Karolinska Institute and QinetiQ
Watch for
MRI finds hidden bone damage
Dysbaric osteonecrosis is the death of bone near a joint, linked to decompression, which can wreck a shoulder or hip. Reviewing the fitness-to-dive checks of 120 professional divers from 2018 to 2024, Körpinar and colleagues found it in 10 (8.3%); magnetic resonance imaging (MRI) found 9 of the 10 and X-rays only 1, and the authors suggest routine MRI for divers with long or intense diving or a history of decompression sickness. A Japanese study using MRI found it in 3 of 62 navy saturation divers.
Achieved whena prospective study in working commercial divers confirms that MRI finds lesions X-rays miss, and a fitness standard for commercial divers adds MRI to periodic checks.
Claimed by
Körpinar and colleagues, and Miyoshi and colleagues, Undersea & Hyperbaric Medicine, 2026
Bubble sounds graded by machine
Trained listeners grade Doppler bubble sounds on the 0 to IV Spencer scale. The group that released the 10,099 recordings says they have allowed a deep learning model that grades bubble loads without a human operator. An earlier version, tested on 274 real recordings, agreed with human graders with weighted kappa scores of 0.74 and 0.69 at two listening sites, which its authors call similar to the agreement between human graders.
Achieved whenan outside group scores the released model on recordings from another laboratory that it was not trained on and reports its agreement with two trained human graders.
Claimed by
Blogg, Azarang and colleagues, March 2026
Open
01
Score tables against bubbles
A public scoring of open decompression models against the 2026 Doppler repository: for each dive whose profile can be rebuilt from the metadata, compute what each model predicts and test how well it ranks the bubble grades the divers actually showed.
The briefCredit · unclaimed02
Test AI on dive tables
A public test of how often general AI models read published decompression tables correctly, with questions such as the required stops for a given depth and time, answered from the table and from memory.
The briefCredit · unclaimed
Open 01 · the brief
Score tables against bubbles
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public scoring of open decompression models against the 2026 Doppler repository: for each dive whose profile can be rebuilt from the metadata, compute what each model predicts and test how well it ranks the bubble grades the divers actually showed.
Why it is still open
The repository's authors invite others to use it and warn that its 12 cases of decompression sickness are too few for risk models, which leaves bubble grades as the outcome to test against. Currens, Doolette and Murphy found in 2025, on 1,196 dives, that the share with decompression sickness rose from 0% at bubble grade 0 to 5.7% at grade 4, so grades carry information about risk. GitHub searches in November 2026 found many open model libraries (dive-deco, libbuhlmann, diy-zhl, a DCIEM 1973 transcription, a two-phase bubble model updated in November 2026), but none scored against this set, and no paper doing so was found. The metadata vary: the authors had time to maximum depth for only 15 dive profiles.
Built with what exists
The repository's master spreadsheet and audio files on Zenodo; open model code from GitHub (dive-deco, libbuhlmann, diy-zhl, edelprino/DCIEM); Python with SciPy for rank correlation and scikit-learn for the area under the curve; an analysis plan posted on the Open Science Framework before any results are run.
First step
Download the 1.2 MB master spreadsheet and count the recordings that give depth, bottom time and decompression schedule.
Done looks like
For at least 1,000 recordings with a rebuilt profile, at least four models scored by the rank correlation between predicted stress and peak Spencer grade and by how well they pick out grades III and IV (area under the curve), with all code and the profile rebuilds published.
Open 02 · the brief
Test AI on dive tables
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of how often general AI models read published decompression tables correctly, with questions such as the required stops for a given depth and time, answered from the table and from memory.
Why it is still open
People now ask chatbots about decompression. Divers Alert Network Europe published a warning in August 2026 that chatbots "can be very wrong" and must not be used to calculate decompression, but it cites no measured test, and a web search in November 2026 found no published study of AI models on table lookups. On GitHub, a Claude Code skill for planning technical-diving decompression appeared in August 2026, so such tools are being built.
Built with what exists
The air decompression tables of the US Navy Diving Manual, revision 7, transcribed to CSV and checked by two people; a script that writes questions and answer keys straight from the CSV; current models through their APIs, including fixed-answer versions of each question for models such as Gemini Decision 1 (Note: "A thousand questions, one pass").
First step
Transcribe one page of the air table to CSV and generate the first 100 questions.
Done looks like
1,000 questions with checked answers, each asked of at least five models with and without the table page supplied; the share answered exactly right published per model, with a notice that the results are not for planning dives.