Neural Operator Captures Three Stiffness Regimes in Plate Flow-Induced Vibration
The preprint presents a single neural operator that jointly evolves coarse flow and structural states for a flexible plate across multiple stiffness-dependent vibration regimes. It introduces a differentiable force module to improve gradient flow and enables interpolation to untrained stiffness values. Evidence rests on autoregressive rollouts up to 1000 steps with preserved frequencies and bounded trajectories.
The arXiv preprint describes a stiffness-conditioned neural evolution operator that represents the plate with 101 Lagrangian tokens and couples them to Eulerian flow via bidirectional cross-attention. Training used staged autoregressive rollouts plus symmetry augmentation on nondimensional bending stiffness as the sole conditioning variable. This single operator reproduces principal flow structures, dominant frequencies, and bounded long-term behavior without separate models per regime.
Conventional surface integrals for aerodynamic forces were replaced by a differentiable derivative-moment transformation around the core vortex region using SDF and smoothed Dirac-delta functions. This change mitigates gradient errors from under-resolved near-wall data while preserving end-to-end differentiability for potential optimization loops in energy-harvesting applications.
Related work on neural operators for Navier-Stokes (Li et al., 2021) and reduced-order FSI modeling shows similar gains in rollout stability but rarely conditions on structural parameters or enforces force consistency via DMT. The current approach therefore bridges data-driven surrogates with physics-based force reconstruction more tightly than prior efforts.
Next steps include extension to 3D geometries and experimental validation; if error remains below 8 % on unseen stiffnesses in wind-tunnel tests within two years, the framework could accelerate parametric studies for flow-energy harvesters.
Chen et al.: Wind-tunnel validation of the operator on a physical plate will show frequency error below 7 % for two untrained stiffness values by Q3 2027.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2609.26816)
- [2]Supporting Source(https://arxiv.org/abs/2010.08895)