Machine learning is used in ship design research mainly as a fast stand-in for slow hydrodynamic calculations, so that thousands of hull shapes can be screened for drag before any is tested in detail. A 2023 Massachusetts Institute of Technology (MIT) study built a public dataset of 30,000 hull designs and used a neural network to cut the calculated drag of one container-ship hull by about 60 percent, but the authors state that the result applies to a simplified model and not to a buildable vessel.
Key takeaways
- The MIT dataset, called SHIP-D, holds 30,000 hulls, each with 32 wave drag coefficients computed for different operating conditions.
- A neural network trained on it predicted wave drag with an R-squared of 0.969, and a genetic algorithm used it to find a hull with 60.7 percent lower total drag under stated conditions.
- The authors report that some optimized hulls are not practical, including an unconstrained case that produced a 685-meter hull.
- A follow-up diffusion model, ShipGen, generated hulls that covered 99.5 percent of the training data's range of designs.
- Regulation is one driver: the International Maritime Organization (IMO) has required efficiency ratings for existing ships since January 1, 2023.
Why hull shape matters and why design is slow
A ship's hull determines how much water resistance the engine must overcome, and resistance drives fuel use. Naval architects normally compare candidate shapes with physics-based simulation and model tests, which are accurate but time consuming. The MIT authors describe ship design as a process with years-long cycles and small production batches, and argue that machine learning could shorten the search if enough shared data existed. Their stated gap was the lack of a publicly available ship design dataset, which the SHIP-D paper by Noah Bagazinski and Faez Ahmed set out to fill. It was presented at the ASME International Design Engineering Technical Conferences in 2023.
What the SHIP-D study actually did
The dataset has three subsets of 10,000 hulls. Each hull is defined by 45 parameters and comes with a mesh, a point cloud, five images, and 32 wave drag coefficients. A residual neural network learned to predict those coefficients, and because the model is differentiable and fast, it could be paired with a genetic algorithm that searched for lower-drag shapes.
In the headline case, a container-ship hull was re-proportioned to 200 meters in length overall and evaluated at 25 knots and a 12.5-meter draft. Optimizing the bow, stern, and bulb while holding the midship body fixed lowered total drag by 60.7 percent in the model. Optimizing the bulb alone gave 54.3 percent. Drag was computed with linear wave theory plus a standard skin friction formula, with viscous pressure drag assumed negligible, so the figure describes a calculation and not a sea trial.
Limits the authors state
The paper lists several constraints on the result. The parameterization cannot represent stepped hulls, strakes, or the flared bow of some naval ships, and the method assumes slender ships. Many hulls in the dataset have high drag, so random sampling rarely finds strong designs. In an unconstrained run aimed at a fixed displacement, the optimizer produced a 685-meter hull, which the authors call impractical, noting that structure and port size impose limits beyond drag. A drag number alone also says nothing about stability, cargo capacity, or cost.
Generating designs instead of ranking them
The same authors later published ShipGen in the Journal of Marine Science and Engineering in 2023. It uses a denoising diffusion model, a type of generative AI, to produce new parametric hull designs under multiple objectives and constraints. With classifier guidance, the model reproduced 99.5 percent of the coverage of the training set's design space, and guided hulls showed an average 91.4 percent reduction in wave drag coefficients compared with the training data mean. The authors present this as a way to speed up early design, with the generated hulls passed on for further analysis.
The rule that gives hull efficiency a price
Efficiency in shipping is partly a regulatory matter. According to the IMO page on ship energy efficiency, the Energy Efficiency Design Index for new ships entered into force on January 1, 2013. The Energy Efficiency Existing Ship Index (EEXI) and the Carbon Intensity Indicator (CII) rating took effect on January 1, 2023. CII rates ships of 5,000 gross tonnage and above from A to E on operational efficiency, and a ship rated D for three consecutive years or rated E must submit a plan of corrective actions.
| Measure | In force since | Applies to |
|---|---|---|
| Energy Efficiency Design Index (EEDI) | January 1, 2013 | New ships of listed types and sizes |
| Energy Efficiency Existing Ship Index (EEXI) | January 1, 2023 | Existing ships of 400 gross tonnage and above |
| Carbon Intensity Indicator (CII) rating | January 1, 2023 | Ships of 5,000 gross tonnage and above |
These rules explain why drag reductions matter commercially, but none of them requires or certifies the use of AI in design.
What to look for in an AI ship design claim
- Is the improvement measured by simulation or by model basin or sea trials?
- Which physics does the underlying calculation leave out?
- Were structure, stability, cargo volume, and port limits part of the optimization?
- Is the dataset public, so another group can repeat the result?
- Is the percentage measured against a real baseline hull or against an average of random designs?
FAQ
Can AI design a complete ship today?
The published research covers hull shape and drag, not a complete vessel. The authors of SHIP-D list structural and port constraints that were not modeled.
Does a 60 percent drag reduction mean 60 percent less fuel?
No. The figure is a calculated reduction in total drag for one hull at one speed and draft. Fuel use also depends on the engine, propeller, weather, and how the ship is operated.
Is the SHIP-D dataset open to other researchers?
The paper presents it as a publicly available dataset that other researchers can use to build and test ship design methods.
Do IMO rules require shipbuilders to use machine learning?
No. The IMO measures set efficiency targets and ratings. How a designer reaches them, by simulation, tank tests, or machine learning tools, is not specified.
Related reading
- Neural networks on ships
- What AI design means in practice
- AI in interior design
- AI and dashboard design
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Bagazinski NJ, Ahmed F, SHIP-D: Ship Hull Dataset for Design Optimization using Machine Learning, ASME IDETC-CIE, 2023 (arXiv:2305.08279)
- Bagazinski NJ, Ahmed F, ShipGen: A Diffusion Model for Parametric Ship Hull Generation with Multiple Objectives and Constraints, Journal of Marine Science and Engineering 11(12), 2023
- Improving the energy efficiency of ships, International Maritime Organization, 2026