Neural operator surrogates predict soft swimmer flows with 3.5% error
AI predicts eel swimmer hydrodynamics at 3.51% error, changing CFD forever
High-fidelity immersed-boundary simulations of soft swimmers like eels are computationally expensive, limiting their use in engineering design, parameter studies, and control. To address this, Mohammad Sadegh Eshaghi and colleagues introduced neural-operator surrogates that predict hydrodynamic fields (velocity, vorticity, pressure) in near real-time. Trained on adaptive fluid-structure simulation data, these models take swimmer geometry and Reynolds number as inputs, bypassing the need to solve the full Navier-Stokes equations.
The planar model outputs two velocity components, scalar vorticity, and pressure simultaneously, achieving a full-domain global relative L² error of just 3.51% on five held-out high-Reynolds-number trajectories. For volumetric eel swimmers, the team used three separate target-specific models sharing a multichannel input. Errors were 3.44% for 3D velocity, 5.58% for vorticity, and 19.2% for pressure—a promising start, though pressure accuracy remains a challenge.
These neural surrogates offer a practical speedup for iterative tasks: instead of rerunning expensive simulations, engineers can instantaneously approximate flow fields. The authors identify pressure accuracy and physical consistency as key areas for future work, likely incorporating physics-informed constraints to enforce conservation laws. The paper is available on arXiv (2608.07722).
- Planar surrogate achieves 3.51% global relative L² error on held-out high-Reynolds trajectories
- Volumetric models hit 3.44% (velocity) and 5.58% (vorticity), with pressure lagging at 19.2%
- Models are conditioned on swimmer geometry and Reynolds number, enabling instant inference
Why It Matters
Fast, accurate flow prediction enables real-time control and design optimization for soft underwater robots.