Science and engineering
Machine learning for ocean modelling
Ocean scientists use machine learning to forecast sea-surface temperature and marine heatwaves, to fill gaps in satellite and float observations, and to emulate or correct parts of numerical ocean models such as subgrid mixing.
How machine learning is used in ocean modelling
Long-lead climate forecasts
Networks trained on climate model output and reanalysis learn ocean patterns, such as sea-surface temperature and heat content anomalies, that come before events like El Niño. They complement dynamical seasonal forecasts.
Real exampleDeep learning for multi-year ENSO forecasts (Ham, Kim and Luo, Nature, 2019)Used a convolutional neural network, trained first on historical climate simulations and then on reanalysis with transfer learning, to make skilful ENSO forecasts up to one and a half years ahead.
Filling gaps in satellite observations
Clouds hide large parts of the ocean in optical and infrared satellite images. Networks trained on the visible pixels reconstruct the missing values, ideally with an error estimate for each one.
Real exampleDINCAE 1.0 (Barth et al., Geoscientific Model Development, 2020)Developed a convolutional auto-encoder that reconstructs sea-surface temperature hidden by clouds in satellite data and provides error estimates for the reconstruction.
Emulating ocean models
An emulator learns to step an ocean model forward from its own output, so long simulations and large ensembles can run far faster. It must stay stable and free of drift over long runs.
Real exampleSamudra ocean emulator (2024)Built an emulator of the ocean component of a climate model that reproduces temperature, salinity, sea surface height and currents through depth, stays stable for centuries and runs 150 times faster than the original model.
Can you learn machine learning for ocean science?
Yes. Ocean scientists can add machine learning to the physics they already know: Python and gridded data tools, statistics and forecast skill scores, classical models, then deep learning for maps and time series.
- PythonLoad NetCDF model output and observations with xarray and plot maps and sections.Learn it in Python
- Data handlingCompute anomalies and climatologies, mask land and align observations with model grids.Learn it in Data Tools
- Statistics and time seriesTrends, autocorrelation and forecast skill scores are the language of ocean prediction.Learn it in Statistics
- Machine learning basicsRegression and tree models for local predictions, validated on separate years.Learn it in Machine Learning
- Deep learning for gridded dataConvolutional networks and U-nets for maps, and sequence models for forecasts.Learn it in Deep Learning
Moving into machine learning for ocean modelling
If you already work in the field
Oceanographers already know the physics, the observing systems and the models. Add PyTorch and work on gap filling or a forecast using satellite or reanalysis data you already use, judged with standard skill scores. Research institutes, met services and ocean forecasting centres are building these teams.
If you come from data science or ML
Machine learning practitioners should learn ocean basics, including circulation, the mixed layer and ENSO, and how satellites and Argo floats make their measurements. NetCDF and xarray are the working tools, and validation must respect time so that future years never appear in training.
Roles that do this workOcean data scientistPhysical oceanographerClimate scientistOcean modeller
Tools you will use
| Tool | What you use it for |
|---|---|
| Xarray | Load, slice and compute on NetCDF ocean grids. |
| Copernicus Marine Toolbox | Download Copernicus ocean reanalysis and observation products from Python. |
| Parcels | Simulate particle trajectories through ocean velocity fields. |
| scikit-learn | Fit regression baselines on lagged ocean variables. |
| PyTorch | Train convolutional and recurrent models on gridded ocean fields. |
Public data to practise on
- NOAA OISST: Daily global sea-surface temperature blending satellite and in-situ observations, suited to forecasting and anomaly studies.
- Argo floats: Temperature and salinity profiles from autonomous floats across the global ocean.
- Copernicus Marine Service: Ocean reanalyses, forecasts and observations; free registration required.
Your first project
Download NOAA OISST daily sea-surface temperature for one region, forecast anomalies seven days ahead with ridge regression on lagged values, and compare it with persistence and climatology using RMSE on a held-out later year.
Common questions
Can machine learning forecast El Niño?
Research models have produced skilful ENSO forecasts more than a year ahead. Forecast centres use such models alongside dynamical forecasts rather than instead of them.
What data is used for ocean machine learning?
Satellite sea-surface temperature, height and colour, Argo float profiles, moorings, ocean reanalyses and output from ocean and climate models. Most of it is openly available, usually as NetCDF files.
Will AI replace ocean models?
Not soon. Emulators are fast but learn from model output and can drift or miss trends outside their training data. They are most useful for fast ensembles and exploring scenarios alongside physics-based models.
Which programming tools do ocean scientists use?
Python with xarray, NumPy and Dask for gridded data, and PyTorch for deep learning. Many ocean models are written in Fortran, which matters if you embed machine learning inside them.
Courses for these skills
- Deep Learning with PyTorchBuild neural and spatiotemporal modelling foundations.Syllabus available · lessons open as produced
- Maths for Machine LearningUnderstand field representations and optimisation objectives.26 published lessons · all planned lessons open
- Statistics and Probability for Machine LearningUnderstand noise and conditional uncertainty.41 published lessons · all planned lessons open
New to programming or maths? Start with Python, or follow the machine learning path in order. Published lessons are free to read; compare Free and Pro.