Machine Learning for Nonorographic Gravity Waves in a Climate Model
Dr. Steven Hardiman
In recent years, the progress of artificial intelligence in generating weather forecasts has been incredibly impressive. How artificial intelligence will be useful on climate time scales is still unclear. In this talk we present one potential application for climate developed at the Met Office. This demonstrates the use of a neural network to mimic the behaviour of one of the sub-grid parameterization schemes used in global climate models: the non-orographic gravity wave scheme. A climate model simulation, using the neural network in place of the existing parameterization scheme, is found to accurately generate a quasi-biennial oscillation of the tropical stratospheric winds, and correctly simulate the non-orographic gravity wave forcing associated with the El Niño–Southern Oscillation and stratospheric polar vortex variability. These internal sources of variability are essential for providing seasonal forecast skill, and the gravity wave forcing associated with them is reproduced by the neural network without explicit training for these patterns.
AI-powered approximation for fast and accurate prediction of nonlinear interactions in wave forecasting
Dr. Jialun Chen
Accurate wave forecasting is crucial for various applications, such as optimising vessel route planning by avoiding rough seas, saving time and fuel consumption, ensuring safe and efficient maintenance periods at offshore wind farms, and designing robust and economical coastal structures. However, this requires an accurate and efficient computation of nonlinear wave–wave interactions, which involve solving a six-dimensional Boltzmann integral. This is an extremely time-consuming process that has challenged researchers for over half a century. To enable practical applications, almost all state-of-the-art operational wave models rely on simplified approximation with known deficiencies. In this study, we present a data-driven approach that could enable fast and accurate predictions of nonlinear wave–wave interaction terms. Our method employs a U-net model that directly maps discrete wave spectra and Discrete Interaction Approximation (DIA) onto their corresponding full nonlinear interaction terms. The model was trained using ERA5 analysis wave spectra spanning various locations and time periods. To assess the model’s robustness, we further evaluate its ability to extrapolate beyond its generalization ability with model integrations. Practical possibilities for further improvements are also suggested and discussed.
AI in Ocean and Coastal Modelling: Research Trends and Editorial Insights
Dr. Bahareh Kamranzad
Artificial Intelligence (AI) is playing an increasingly important role in the monitoring, modelling, and forecasting of ocean and coastal environments. In this talk I will reflect on developments in the field from both a research and editorial perspective, drawing on my experiences in AI-based wave forecasting, as well as involvement with the journals Ocean Engineering, Coastal Engineering, and the recent special issue on “Artificial Intelligence for Ocean Monitoring and Modelling” in Applied Ocean Research. The presentation will explore current trends and recent contributions in the literature, highlight common challenges such as model generalisation and interpretability, and discuss emerging directions that are shaping the future of AI applications in marine science. The aim is to offer an overview of progress in the field and encourage further interdisciplinary collaboration.
Physics-informed machine learning methods for ocean wave data assimilation
Svenja Ehlers
Accurate phase-resolved ocean wave prediction relies on precise spatio-temporal initial conditions. However, the inherent sparsity of wave measurements in space or time necessitates a reconstruction of wave information, known as data assimilation, which poses challenges and demands significant computational resources when using conventional grid-based solvers. To overcome these limitations, we first investigate physics-informed neural networks (PINNs) for wave data assimilation. PINNs integrate sparse observational data with physical laws, in our case by parameterizing solutions to the fully nonlinear potential flow equations as neural networks. This PINN framework enables the reconstruction of both, wave surfaces and the corresponding physically consistent potential field, from sparse surface buoy measurements only. While PINNs achieve accurate reconstructions, they require retraining for each new measurement instance, limiting real-time applicability. To address this, we introduce a Physics-Informed Neural Operator (PINO) [3] framework that generalizes the assimilation across multiple wave instances. Once trained, the PINO provides near-instantaneous reconstructions of wave surface elevation and surface potential and, unlike classical supervised learning, does not require fully resolved ground truth data . Instead, similar to the PINN approach, they also leverage a physics-based loss function to reconstruct wave surfaces between sparse data measurements. Our approach demonstrates the potential to enhance the reliability and efficiency of wave predictions by combining physics with data-driven methods, offering benefits for ocean engineering applications.
Shock wave singularities, phase transitions and learning in random neural networks
Antonio Moro
We illustrate the mean-field analog of the p-star model for homogeneous random networks and compare its behaviour with that of the p-star model and its classical mean-field approximation. We show that the partition function of the mean-field model satisfies a sequence of partial differential equations which admit shock wave solutions. Shocks are associated with phase transitions and stable states, i.e. learned patterns, are selected consistently with the Maxwell construction. As an example, we discuss in detail the 3-star model. Monte Carlo simulations show an excellent agreement between the p-star model and its mean-field analog at the macroscopic level, although significant discrepancies arise when local features are considered.
Swans, swarms and surrogates – operational AI Wave forecasts in the (Dutch) North Sea
Joost den Bieman
In recent years, quite some research and development has been conducted around AI wave prediction in the (Dutch) North Sea by the research institute Deltares, much of it for the Dutch ministry of Infrastructure and Water Management. In this presentation, a few examples will be discussed in-depth, including hybrid wave modelling combining numerical and data-driven models in an operational setting, surrogate wave modelling and short lead time offshore workability forecasting using local measurements from drone swarms.
Machine Learning for Wave Breaking Evolution
Tianning Tang
This talk focuses on how we can use symbolic machine-learning methods to explore the underlying physics of a commonly observed yet not fully understood phenomenon -- the breaking of ocean waves. In our work, we train our model from a dataset generated from expensive numerical methods to explore wave-breaking evolution. Our work discovers a new boundary equation that provides a reduced-order description of how the surface elevation (i.e., the water-air interface) evolves forward in time, including the instances when the wave breaks -- a problem that has defied traditional approaches. Further expert AI collaborative research reveals the physical meaning of each term of the discovered equation, which suggests a new characteristic of breaking waves in deep water -- a decoupling between the water-air interface and the fluid velocities. This novel reduced-order model also provides computationally efficient ways to simulate breaking waves for engineering applications.