Generative AI design cuts fuel use and emissions

Designers make the decisions that will determine around 90% of a vessel’s eventual performance at the concept stage, before a hull form is even fixed. Naval architects have traditionally tackled this with the design spiral, refining a shape through repeated cycles of assessment and adjustment, a process that delivers thoroughness but sacrifices speed, and tends to favour conservative outcomes. Generative AI is showing it can propose hull forms that go far beyond what that process would typically surface. The harder question is whether engineers can trust the unconventional designs, simulate them with enough confidence to predict real performance, and build them without the costly rework that so often follows an ambitious concept into the design office.

Origins of the project

Compute Maritime, a London-based deep tech company, built its NeuralShipper tool to tackle this problem directly, using generative AI to produce novel hull forms as native CAD geometry rather than abstract concepts requiring translation. GenDSOM (short for generative design, simulation, optimisation and manufacturing) set out to prove that NeuralShipper’s outputs could be trusted, simulated to a validated standard and carried through to a buildable vessel, replacing the traditional design spiral with a single, intelligence-led loop.

It ran as a seven-month, roughly £700,000 feasibility study under Round 6 of the Clean Maritime Demonstration Competition, part of the Department for Transport’s UK SHORE programme, delivered by Innovate UK. From September 2025, the consortium took BYD Naval Architects’ B-WIND crew transfer vessel (CTV) from brief to validated design. The result is a 32.5m twin-hull aluminium CTV, carrying 24 technicians and four crew at 25knots. Compared with a like-for-like diesel baseline, it saves 101,671litres of fuel and 258.7tonnes of CO₂ per vessel, per year.

For the engineering, it had to be performant, predictable and producible. Lose any one, and the output is a rendering, not a vessel. NeuralShipper generates hulls as high-fidelity NURBS surfaces rather than mesh, since