A masked Fourier neural operator for fast simulation of vortex-induced vibration

Abstract

Achieving fast simulation of fluid–structure interaction (FSI) is a longstanding challenge in computational mechanics, which plays an important role in system identification and fast control. This work introduces a masked Fourier neural operator (MFNO) as a fast and accurate surrogate for two-way FSI in planar vortex-induced vibration (VIV) of a rigid circular geometry. The framework replaces the fluid solver with a neural operator-based model and integrates a mask operation that embeds structural displacement to enable a clear separation at the fluid–structure interface. This design mitigates error accumulation during rollout predictions and preserves the flexibility of the Fourier neural operator (FNO) for direct inference on refined grids without retraining under the same CFD settings, producing smoother flow-field predictions. The performance of the MFNO is examined on a VIV benchmark using weakly and strongly coupled schemes. Results demonstrate that the strongly coupled MFNO achieves more accurate predictions than the weakly coupled MFNO and FNO-based strategy, while providing an approximately 158-fold speedup over CFD–CSD simulations. The proposed model performs well on unseen reduced velocity cases, including interpolation and nearby extrapolation conditions, and enables open-loop rollout without ground-truth corrections for more than 8 vibration cycles.