AI boosts understanding of ocean dynamics and marine structure safety
image: Machine learning framework for reduced-order modeling and control in fluid–structure interaction. This figure illustrates a machine learning framework that compresses high-dimensional fluid–structure interaction (FSI) data into a low-dimensional latent space using an encoder–decoder architecture. Temporal evolution in the latent space enables rapid prediction of flow fields, while a reinforcement learning agent interacts with the environment through local measurements to determine control actions. The model outputs predicted FSI states efficiently, providing a data-driven approach for real-time simulation, flow control, and decision-making in complex FSI systems.
Credit: Ocean, Tsinghua University Press
Fluid–structure interaction (FSI) plays a foundational role in offshore engineering, influencing the stability, fatigue life, and performance of marine risers, floating wind turbines, subsea pipelines, fish cages, propellers, and hydrokinetic energy devices. When fluid forces deform structures, the resulting motions alter flow patterns in return, forming a tightly coupled nonlinear system. High-fidelity simulations using Navier–Stokes solvers demand massive central processing unit (CPU) hours, while experiments struggle with sparse sensing and moving-body limitations. As ocean renewable energy expands and system scales grow, faster, data-efficient and generalizable modeling techniques are urgently needed. Due to these challenges, leveraging machine learning (ML) to accelerate FSI modeling, prediction, and control has become necessary for further development.
A recent review conducted by the University of Stavanger (Norway) presents a comprehensive overview of how ML is accelerating progress in FSI research. Published (DOI: 10.26599/OCEAN.2025.9470002) in Ocean in 2025, the review highlights advances in data-driven flow analysis, reduced-order modeling, and intelligent control strategies for marine structures. By integrating computational physics with ML, the review showcases emerging pathways to enhance prediction accuracy, reduce simulation cost, and support innovation in ocean renewable energy engineering.
The review divides ML-enabled FSI progress into three major directions—feature detection, dynamics prediction, and flow-structure control. Unsupervised and supervised learning techniques, including proper orthogonal decomposition (POD), dynamic mode decomposition (DMD), convolutional neural network (CNN) autoencoders, variational autoencoders, generative adversarial networks (GANs), and sparse identification of nonlinear dynamics (SINDy), have successfully extracted low-dimensional modes from turbulent wake flows, revealing coherent vortex structures behind cylinders and hydrofoils. These models can reconstruct flow fields from limited inputs, providing high-resolution representations at a fraction of simulation cost. Recurrent architectures such as long short-term memory network (LSTM) and transformers have improved temporal forecasting of vortex shedding, while reduced-order models enable efficient load estimation and structural response prediction.
Beyond analysis, reinforcement learning is beginning to achieve real-time control, suppressing vortex-induced vibration or enhancing wave-energy harvesting in numerical tests. Physics-informed neural networks further integrate Navier–Stokes equations directly into training, enabling solutions without mesh generation and reducing reliance on extensive datasets.
The authors identify challenges ahead—nonlinear energy transfer in turbulence, sparse experimental sensing, realistic structural geometries, and high-Reynolds-number conditions. Future success will require hybrid modeling that respects physical laws while utilizing ML efficiency, especially for real-ocean deployments.
“Machine learning is not replacing classical FSI methods—it is expanding what we can solve,” the authors note. “The ability to decode flow physics from data, predict future states, and operate controllers adaptively offers new pathways for renewable energy systems and offshore infrastructure. As models become more physically informed, ML could transform ocean engineering in the same way it transformed vision and language processing.”
By integrating ML into FSI workflows, engineers may accelerate design cycles for tidal turbines, extend fatigue life of subsea pipelines, develop self-optimizing risers, and achieve real-time control for energy harvesting devices. ML-assisted reduced-order models could reduce CPU demand from thousands of hours to seconds, enabling fast risk assessment during storms or installation operations. In the long term, bridging mathematical modeling, big-data ocean sensing, and physics-informed algorithms may lead to digital twins of marine structures and autonomous offshore systems. The review suggests that interdisciplinary innovation will be key to transitioning FSI-ML research from laboratory cases to real-ocean industry deployment.
About Ocean
Ocean is an international, peer-reviewed, open-access journal that provides a multidisciplinary platform for cutting-edge research and practical applications in the fields of ocean science, marine technology, and marine engineering. The journal publishes articles, reviews, and perspectives aimed at advancing theoretical, numerical, site-based, and experimental developments to promote global sustainability.