Data-driven reduced-order modeling for undulating swimmers

Bioinspired underwater robots have garnered significant academic interest due to their exceptional maneuverability, propulsion efficiency, and deceptive appearance. However, designing a system that accurately mimics the fluid muscle movements of fish poses significant challenges, particularly in experimental modeling due to limited experimental data and high costs in physical environments. Numerical simulation offers a viable solution, but current methods are computationally expensive. Recent advances in computing technologies enable nonintrusive modeling approaches and machine learning to build efficient and effective surrogate models. This study presents a data-driven reduced-order modeling (ROM) approach, combining proper orthogonal decomposition with linear multistep neural networks, for long-term prediction of fluid dynamics around undulating swimmers in the thrust-producing range Re=5000 and St=0.3−0.7. The results are compared with direct numerical simulation data obtained by solving the Navier–Stokes equations. Our nonintrusive approach achieves high accuracy, addresses computational challenges, and paves the way for innovative designs and strategies in natural and artificial swimming mechanisms. The effectiveness of the proposed ROM across various undulation frequencies highlights its potential for advancing bioinspired underwater robotics, optimizing adaptive propulsion systems, and enabling future real-time control applications.