GB1 uses two kinds of AI to explore new yacht hulls

GB1 and PhysicsX announced their new hull-design workflow on 21 October. Large Geometry Models propose shapes; Large Physics Models predict performance and guide further exploration through active learning. The new work extends a partnership begun for the 2027 campaign beyond optimisation within a predefined design space.[1]

GB1 reports a five-fold reduction in high-fidelity fluid-flow simulation requirements in early tests. The announcement does not publish the underlying test protocol.[1]

A faster search needs more shapes to consider

An August technical article by PhysicsX researchers explains the geometry problem. Parametric computer-aided design lets engineers vary the dimensions and relationships specified in advance. A generative model can instead learn a family of shapes and propose new members. The useful change is in what can be considered, as well as how quickly a candidate can be evaluated.[2]

Their published examples adapt a pretrained 3D model to engineering data, then repair its output into meshes suitable for numerical solvers. They describe steering generation with a learned physics predictor or with simulation itself. These are broader method examples, rather than a disclosure of GB1’s exact implementation.[2]

The design problem has many operating states

PhysicsX’s July account of the partnership describes models trained on simulation and experimental data, and an engineering workflow that generates data, trains models and evaluates designs. GB1’s head of design, Nick Holroyd, points to the many combinations of states in which the boat can sail. That makes the search more demanding than finding one appealing shape for one condition.[3]

A design team therefore needs a way to ask a sequence of questions about a candidate. Which circumstances suit it? Where do the trade-offs appear? What should be investigated next? These are engineering decisions that a faster prediction makes easier to revisit; they remain decisions about a physical object and its intended use.[3]

NVIDIA’s account adds a useful detail: the partners use Omniverse libraries to visualise hull geometry and pressure fields. Pairing a proposed shape with its predicted field gives engineers something to inspect together. They can connect a change in the result to the part of the geometry that changed, keeping a design review tied to the shape under consideration.[4]

Report an error or suggest a correction