Learning the Fidelity Gap: Physics-Informed Graph AI for Marine Propeller Design
Literature Highlight
Everybody doing more simulation wants higher fidelity - more accurate solver, finer mesh resolution to pickup on details better, etc. ML surrogates have been show to help map between the simulations they can run (coarse, moderate fidelity), to fidelity/resolution of results that are prohibitively expensive (too big or too slow to run simulations on).
Designing a marine propeller is a balancing act between performance, efficiency, noise, vibration, and durability. At the heart of that challenge is a deceptively complex question: what is happening to pressure across the blade surface as the propeller moves through a non-uniform wake?
Those pressure distributions are central to understanding blade loading and cavitation behaviour. Cavitation, the formation and collapse of vapour cavities in low-pressure regions, can affect efficiency, generate noise and vibration, and contribute to erosion. For designers, being able to predict cavitating pressure fields accurately is therefore essential. But there is a practical obstacle: the simulations that capture these effects with the highest fidelity are often too computationally expensive to use at the scale required for broad design exploration and optimization.
This work by Hubbard et al (link here) tackles that bottleneck with a physics-informed graph learning approach that bridges two gaps at once: the gap between non-cavitating and cavitating physics, and the gap between coarse and fine simulation meshes.
“Fidelity Scaling via Physics-Informed Graph Learning for Marine Propeller Blades”, by Ian Hubbard*, Evert-Jan Foeth.
Rather than replacing physics-based simulation outright, the method learns how to scale from a cheaper, lower-fidelity solution to a more detailed cavitating pressure field. In the study, paired simulations are generated using the boundary element method solver PROCAL. The input is a coarse, non-cavitating blade-surface pressure field, while the target is a finer-resolution cavitating pressure field. The model learns the difference between them, effectively turning a fast wetted-flow calculation into a high-value prior for predicting cavitation-sensitive pressure distributions.
The core idea is elegant: use what the physics solver already knows, then let machine learning learn the correction.
To do this, the blade surface is represented as a graph. Each point on the coarse blade mesh becomes a node, with connectivity derived from the structured panel grid. This allows a graph neural network encoder to learn local and global patterns across the blade surface. But the real challenge is that the model must transfer information from a coarse mesh to a finer one. Simple interpolation can be brittle, especially on a complex blade geometry where points may be close in 3D space but far apart along the actual surface.
The authors solve this by introducing a shared UV parameterization of the blade. The structured blade grid is unfolded into a two-dimensional surface coordinate system, giving both the coarse and fine meshes a common reference frame. This surface-aware representation avoids misleading shortcuts through the blade thickness and provides a natural way to map information between resolutions.
On top of that UV mapping, the model uses a cross-attention decoder. Unlike standard transformer attention, where attention weights are learned from query and key projections, this architecture uses a geometric distance kernel in UV space. Fine-mesh query points attend to coarse encoder nodes according to their distance along the blade’s surface parameterization. The result is a differentiable, geometry-aware coarse-to-fine transfer mechanism that replaces heuristic interpolation with an end-to-end learning component.
This matters because the decoder is not tied to a fixed output mesh. Since it can be queried at arbitrary UV locations, the approach has a degree of output-side discretization independence. In practical terms, the model can learn from structured panel-mesh data while still producing predictions on a finer grid in a principled way.
The scale of the dataset is also notable. The study uses approximately 366,000 graph samples spanning thousands of propeller variants, multiple operating conditions, cavitation numbers, wake fields, and blade positions over a full revolution. Each snapshot captures a different instantaneous inflow condition caused by the non-uniform ship wake, making the dataset rich in both geometric and operational variation.
A key part of the work is the comparison between two models. The first, Model-A, is a geometry-only baseline. It receives blade geometry and operating-condition information but no low-fidelity pressure field. The second, Model-B, receives the same information plus the coarse wetted pressure coefficient as a per-node input feature.
The difference is significant. The physics-informed model consistently outperforms the geometry-only baseline across mean squared error, mean absolute error, and R². More importantly, it is more data-efficient. Because the wetted pressure field already contains much of the spatial structure of the final cavitating pressure distribution, the model does not need to learn the entire pressure field from scratch. Instead, it learns the cavitation-driven perturbation. That makes the learning task easier, more physically grounded, and more accurate with less data.
The holdout results illustrate this clearly. On unseen propeller cases, the physics-informed model more closely follows the target cavitating pressure distributions at multiple radial blade sections. It performs especially well in sensitive regions such as pressure peaks and low-pressure leading-edge zones, where cavitation effects are most important. Surface plots show that Model-B better preserves localized low-pressure structures, while the geometry-only model tends to smooth them out.
The study also evaluates integrated quantities such as thrust coefficient, computed directly by numerically integrating the predicted pressure fields. This is an important test because engineering design decisions often depend not only on local field accuracy but also on whether the predicted field produces the right global performance metrics. Here again, the physics-informed model shows a clear improvement over the baseline, demonstrating that better pressure-field prediction translates into better integral performance estimates.
For marine propeller design, this points toward a powerful workflow. Designers could use lower-cost simulations to explore large design spaces, while a trained graph model rapidly estimates cavitation-aware high-resolution pressure fields. That could make it easier to screen candidates, identify promising geometries, and reserve expensive high-fidelity simulations for the most critical cases.
If you like this content, feel free to join my community!
If you want to build basic fundamentals and competency at this intersection of AI/ML and engineering, checkout the online curriculum I have created: