By Matthieu Caillaud · Founder, oceanographer
The coastal zone is one of the hardest to forecast: waves, tidal currents, local wind and bathymetry interact over a few hundred metres. No single operational model describes all of these processes on its own. The answer is not one more model, but the reasoned combination of several sources, refined by deep learning.
Three models, three complementary roles
Each data source answers a specific question. Confusing them leads to wrong forecasts; articulating them covers the whole chain from the open sea to the coast.
- WaveWatch III (WW3) — the sea state: significant wave height, period and direction at the basin scale, forced by wind.
- CROCO — high-resolution coastal circulation: currents, water level, temperature and salinity, down to a few tens of metres near the shoreline.
- Copernicus Marine (CMEMS) — the basin-scale reanalyses and forecasts that provide the boundary conditions and a validation reference.
In practice, CMEMS and WW3 feed the boundaries of a CROCO domain nested over your area of interest. The result is a physically consistent description, from the open sea to the beach.
Why AI comes in here
Increasing the resolution of a circulation model has a cost. For many operational needs — alerting, routing, operating a site — this cost is neither sustainable nor necessary on a continuous basis.
On temporal emulation, a Fourier neural operator (FNO) trained on a 250 m CROCO ground truth evaluates the barotropic currents of the Brest roadstead over multi-start time windows. The experiment requires an initial state supplied by the physical model; from a harmonic initial condition, it is not an autonomous forecast.
Downscaling proper is a distinct problem, and that work is under way: we will train convolutional neural networks (UNet architectures) to reproduce the relationship between a low-resolution field and its high-resolution version, learned on a set of reference CROCO simulations. The specification is settled; the measured figures will come once the model has been trained and confronted with ADCP data. We will not publish a gain before we have measured it.
Validate, always
A forecast is only worth what it is measured against. We systematically quantify the model's error relative to the available observations — buoys, HF radars, AWAC sensors, satellite altimetry — through explicit metrics: RMSE, bias and correlation, variable by variable.
This step is not an end-of-project detail. It is what turns a model output into information you can decide on, and what defines the domain of confidence within which the forecast can be used.
What it changes for you
This hybrid approach — physical models for consistency, AI for resolution and speed, validation for confidence — makes it possible to deliver usable coastal forecasts without heavy computing infrastructure. It applies to alerting (flooding, harbour agitation), metocean weather routing, and site characterisation (marine energy, structures, aquaculture).
Do you have a forecasting need on a specific coastal area? Let's talk about your project — we always start from a diagnosis of the available data and the error acceptable for your use.
Sources and references
The metrics, comparisons and maps specific to this article are OceanData Consulting results and calculations; the links below document the datasets, standards and publications used.
Facing a similar modeling or metocean data challenge?
From physical model calibration to operational AI post-processing, we help turn marine observations into validated, decision-ready forecasts.