Industries / Transport and vehicles / Shipbuilding and repair
Design
Naval architects decide a ship's shape, size, structure and layout, and check by calculation, model tests and computer simulation that it will float upright, reach its speed and hold together before any steel is cut.
6
Milestones achieved, from 1628 to 2023.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1628
Vasa capsizes in port
The Swedish warship Vasa sank in Stockholm harbor on 10 August 1628, minutes after setting sail for the first time; the Vasa Museum says all but 30 of the crew and guests survived. Her upperworks were too tall and heavy for the hull below the waterline, and the museum says her designer "had no way to calculate a proposed ship's performance in advance," because the mathematics of stability was more than a century away. Before she sailed, thirty men ran back and forth across the deck to make her roll; she heeled so far after a few runs that the test was stopped, but the king was not told. The inquest that followed punished no one.
LessonVasa's own stability test showed the danger, and a test only helps if its result can stop the launch.
Credit
Henrik Hybertsson (designer) and Hein Jacobsson (builder); account by the Vasa Museum
1746
Stability by calculation
Pierre Bouguer's Traité du navire (1746) defined the metacenter, a point whose height above a ship's center of gravity shows whether the hull will come back upright after it heels; Leonhard Euler reached an equivalent test in Scientia Navalis (1749). In 1748 a junior constructor at Brest, Clairin-Deslauriers, ran what historians Horst Nowacki and Larrie Ferreiro call the first recorded inclining experiment (heeling a finished ship with known weights to measure its stability) on the 74-gun Intrépide. They credit the Swedish designer Fredrik Henrik af Chapman, whose treatise appeared in 1775, with making the metacenter calculation "a routine matter among naval architects."
Credit
Pierre Bouguer and Leonhard Euler
1871
Ship models in a tank
William Froude, who had towed small models since 1863, persuaded the British Admiralty to pay for a towing tank at Torquay, England: about 90 m long and 12 m wide, with a powered carriage that pulled scale models while their resistance was measured. Froude split a hull's resistance into a friction part and a residual part (mostly from making waves), each scaled to full size by its own rule, and model testing still rests on that split. M. W. C. Oosterveld of MARIN, the Dutch maritime research institute, dates the tank to 1871 and counts about 200 towing tanks in operation when he wrote, after 1987; tank staff have met as the International Towing Tank Conference (ITTC) since 1933.
Credit
William Froude; British Admiralty
2013
MOL Comfort breaks in two
On 17 June 2013 the 8,110 TEU container ship MOL Comfort (TEU: twenty-foot container units), built by Mitsubishi Heavy Industries, cracked across its middle in the Indian Ocean; the two halves drifted and sank, and no one died. The ship had passed every applicable rule. Japan's Committee on Large Container Ship Safety reported in March 2015 that, by simulation, the hull could have failed, because the rules did not count water pressure on the bottom plating together with hull bending, nor whipping (the hull vibrating after the bow slams into waves); inspections of sister ships had found bottom plates buckled by about 20 mm. It asked for the rules to change, and IACS, the association of classification societies (the private bodies that write hull rules and check ships against them), wrote two new requirements: UR S11A on container ship hull strength and UR S34 on computer stress analysis.
Lessonmeeting the rules is not the same as being strong enough, and the rules for big container ships were rewritten only after one broke.
Credit
Committee on Large Container Ship Safety (Japan's Ministry of Land, Infrastructure, Transport and Tourism); IACS
2021
Computers stand in for tanks
In June 2021 the International Maritime Organization (IMO) adopted guidelines for certifying a ship's EEXI (Energy Efficiency Existing Ship Index, a carbon rating for ships already in service). They let numerical calculations, in practice computational fluid dynamics (CFD: computer simulation of the water flowing around the hull), replace model tests if the method has been validated against sea trials or model tests of a parent hull and the verifier approves; they point to ITTC procedures 7.5-03-01-02 and 7.5-03-01-04 as quality standards.
Credit
IMO Marine Environment Protection Committee (resolution MEPC.334(76)); ITTC
2023
Approval from a 3D model
Designers have long sent classification societies 2D drawings for approval, then copied the society's remarks back into their 3D models by hand. In January 2023 Cyprus Shipping News reported that Bureau Veritas had reviewed the complete 3D steel hull model of a Damen 2,500 m³ trailing suction hopper dredger, sent in OCX (Open Class 3D Exchange, an open file format for ship structure models built by software firms and classification societies) through NAPA's software. The report called it "the first complete steel hull to achieve class approval by 3D design," a step toward approval in principle.
Credit
Damen Shipyards (Damen Engineering Gdansk), Bureau Veritas and NAPA
Watch for
A rules assistant that cites
DNV says RuleAgent, its AI tool, answers plain-language questions across more than 30,000 pages of DNV rules by returning the relevant rule paragraphs, short summaries and links to the official text so users can check them.
Achieved whenan outside group scores RuleAgent's answers on a public question set and publishes how often the answer and the cited paragraph are right.
Claimed by
DNV, March 2026
Approval on the 3D model
ClassNK released guidelines for 3D Model-Based Approval and says it has "commenced operation" of a workflow in which OCX 3D models, not 2D drawings, are the primary basis for approval, after a pilot with 10 Japanese companies.
Achieved whena shipyard or owner, not only ClassNK, reports a delivered ship whose hull structure was approved on the 3D model as the main record.
Claimed by
ClassNK, May 2026
AI places hull supports
Korean Register (KR) and HD Hyundai Samho say a one-year joint project produced an AI model that works out the number and position of hull supports "in near real time," a job that used to need repeated structural analysis, plus an in-house language model for searching shipyard design documents.
Achieved whenresults from real ships, with the AI's support layouts checked against full structural analysis, are published or confirmed by a third party.
Claimed by
KR and HD Hyundai Samho, June 2026
Open
01
Testing AI on class rules
a public set of about 300 questions taken from the IACS Unified Requirements (the minimum rules all member classification societies adopt), each with the correct answer and the clause it comes from, and a script that scores any AI assistant on correct answers and correct citations.
The briefCredit · unclaimed02
Hull AI against tank data
a test that runs hull-drag predictors trained on synthetic data, starting with the model released with MIT's Ship-D dataset, on public benchmark hulls that have measured towing-tank results, and publishes predicted against measured resistance.
The briefCredit · unclaimed
Open 01 · the brief
Testing AI on class rules
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
a public set of about 300 questions taken from the IACS Unified Requirements (the minimum rules all member classification societies adopt), each with the correct answer and the clause it comes from, and a script that scores any AI assistant on correct answers and correct citations.
Why it is still open
Rule assistants are now offered (DNV's RuleAgent, above), but searches of the web and arXiv in November 2026 found no public test of AI answers on classification rules. The nearest test, DGEval (August 2026), scores AI models on 1,678 questions about the IMDG Code, the rules for carrying dangerous goods by sea, built partly from a commercial e-learning platform; it does not touch hull or class rules. MIT's MiDShip dataset of 12,753 synthetic cargo-hold designs checked against 25 constraints from ABS rules (2026), says data-driven design is limited by "the lack of structured datasets linking design geometry, structural performance, and rule-based constraints," and it tests designs, not answers to questions.
Built with what exists
the IACS Unified Requirements, Python, general AI models such as Claude Fable 5.2 (see the Note "Three working devices") and open-weight models, and DNV RuleAgent where access allows.
First step
write 30 questions with clause references from one requirement, such as UR S11A on container ship hull strength, and try them on two models.
Done looks like
300 questions whose answers two people with naval architecture training have checked, and published scores for at least three assistants giving the share of answers that are right and cite the right clause.
Open 02 · the brief
Hull AI against tank data
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
a test that runs hull-drag predictors trained on synthetic data, starting with the model released with MIT's Ship-D dataset, on public benchmark hulls that have measured towing-tank results, and publishes predicted against measured resistance.
Why it is still open
Ship-D's 30,000 hulls each carry 32 wave-drag coefficients computed by a formula (the Michell integral), not measured. Its authors checked that formula against towing-tank data for one naval hull from the David Taylor Model Basin at moderate speeds, and used 12 public hulls only to show that their shape parameters can rebuild them; an arXiv search in November 2026 found no follow-up comparing its predictions with tank results. Measured data exist for benchmark hulls such as the KCS container ship, which the Korea Research Institute for Ships and Ocean Engineering (KRISO) tank-tested so that computer methods could be checked against it.
Built with what exists
Ship-D's data and code, hull geometry and resistance data from the Tokyo 2015 CFD workshop and SIMMAN workshop sites, the ITTC-1957 friction line to add friction drag, the Holtrop-Mennen empirical method as a baseline, and Python.
First step
download the KCS geometry and test conditions from the Tokyo 2015 workshop site and fit the hull with Ship-D's parameters.
Done looks like
a table for at least three benchmark hulls at their test speeds, giving each predictor's error in total resistance in percent, next to the error of the Holtrop-Mennen baseline.