Geometric deep learning surrogate for hydromechanics

Problem

Early hull design is a numbers game. A naval architect wants to try many hull variants and see which behaves best in the water, but each hydromechanics evaluation runs as a full simulation that takes hours. That is fine when you are validating one final design. It is useless when someone wants to compare twenty options before committing to a direction.

The result is that exploration barely happens. Designers pick a small number of candidates they already believe in, evaluate those, and move on. The simulation that should inform the design instead only confirms it, late, and at a compute cost that adds up fast across a project.

Approach

I built a surrogate: a machine learning model that learns the relationship between a hull’s shape and its hydromechanic response, then predicts that response directly, without running the simulation.

The decision that mattered was to work on the geometry itself rather than on a handful of hand-picked parameters. A parametric surrogate only knows the sliders you gave it and breaks the moment a hull falls outside that box. A geometric deep learning model operates on the hull mesh, so it generalises across shapes instead of interpolating between a fixed set of knobs. I built it on a foundation model and trained the whole pipeline in Databricks, with MLflow tracking runs so the experiments stayed reproducible as the dataset grew.

Result

For the same evaluation, the surrogate returns an answer roughly 10× faster than the existing simulation pipeline, and at materially lower compute cost. In practice that changes the workflow rather than just the timing: exploring many variants in early design becomes cheap enough to actually do, so the simulation starts informing decisions instead of rubber-stamping them.

The honest boundary: the surrogate is an accelerator for exploration, not a replacement for validation. Final designs still go through full simulation. The model earns its place by making the first ninety percent of the search fast, not by being trusted with the last ten.

What I’d do differently

I underestimated how much of the work was data, not modelling. The mesh preprocessing and the curation of a clean, consistent training set took far longer than getting the network to train, and I built that pipeline reactively instead of designing it up front. Next time I would treat the geometry data pipeline as the first-class deliverable it turned out to be, and stand it up before touching the model.