Reduced-Order Modeling of Flow Around Cylinders Using Machine Learning

As global climate change, ocean pollution, and sea-level rise continue to impact marine environments, understanding fluid flow behavior is increasingly critical in ocean engineering. While partial differential equations (PDEs) accurately describe fluid dynamics, traditional numerical methods such as finite difference, finite volume, and finite element methods are computationally expensive due to high-resolution requirements. Reduced Order Modeling (ROM) method has emerged as an efficient alternative, simplifying complex flow fields without significantly sacrificing accuracy. This study introduces two novel ROM methods by coupling Proper Orthogonal Decomposition (POD), Radial Basis Function Neural Networks (RBFNNs), Autoencoders (AEs), and Fully Connected Neural Networks (FCNNs): the AE-FCNN ROM method and the AE-FCNN-POD-RBFNN (AFPR) ROM method. To validate these methods, a series of numerical experiments involving flow around a cylinder were conducted. The results demonstrate that both methods can effectively reconstruct the flow field, with the AFPR ROM method outperforming the AE-FCNN ROM method in terms of accuracy. This research highlights the potential of hybrid ROM methods in ocean engineering, providing new solutions for efficient and accurate simulations of fluid flows, pollutant transport, and other complex dynamics.