Developing machine learning (ML)-based Earth system model components is central to the AI activities implemented by ECMWF and their partners within the Destination Earth (DestinE) initiative of the European Commission. This blog series shares insights from experts building these components, highlighting their approaches and contributions towards a coupled ML-based Earth system model. 

Learning the evolution of the world’s oceans 

Oceans are essential to life on Earth. They play a key role in regulating the planet’s climate and weather systems by redistributing the sun’s heat through ocean currents. On climate timescales, the deep ocean acts as a major sink of both heat and carbon, influencing how the climate evolves. 

Oceans are also vital for global economies and human well-being. They enable international shipping, provide natural resources including fish, minerals, and energy, and support recreation and tourism. In addition, they sustain an incredible diversity of life, from microscopic plankton to large marine mammals. Marine organisms play a crucial role in producing oxygen—phytoplankton alone are responsible for a significant share of the oxygen we breathe. 

Predicting ocean conditions over various timescales allows us to better understand and manage this precious resource and enables us to make the best use of the ocean, i.e. through ship routing. The ocean and atmosphere are closely linked, and improving ocean predictions contributes to more reliable weather predictions, improved early warnings for extreme events, and better protection for communities and economies.  

From physics-based models to machine learning 

To make these predictions, scientists have traditionally relied on physics-based ocean models. These models simulate the movement of water, heat, and salt by solving equations that describe the underlying physical processes. While skilful, these are computationally expensive and can be challenging to run at high resolution. 

In recent years, advances in machine learning (ML) have opened up new possibilities. By learning directly from data, ML models can complement traditional approaches, offering faster predictions while maintaining a high level of accuracy. This creates an opportunity to rethink how we model the ocean. 

At ECMWF, we are developing an ML-based ocean model to enable faster and more accurate predictions across a range of timescales. ML-Ocean is a data-driven global ocean model developed within the Destination Earth (DestinE) initiative of the European Commission. By leveraging state-of-the-art machine learning techniques and high-quality reanalysis data from ECMWF’s ocean reanalysis system (ORAS6), it provides an accurate and computationally efficient representation of the ocean system. 

The world’s oceans are a particularly difficult target for machine learning. Ocean behaviour varies dramatically, from very fast-moving surface currents to the slow evolution of deep-ocean temperatures. At the same time, the global ocean contains a variety of dynamic environments, from highly energetic boundary currents such as the Gulf Stream to vast, relatively calm interior regions. A successful model must learn to represent all of these behaviours simultaneously. Adding to the challenge, intricate coastlines and seafloor features shape ocean circulation in complex ways, making it even harder to build models that perform reliably across the globe. 

How ML-Ocean learns 

Like all ML models, ML-Ocean identifies patterns in how the ocean evolves over time and makes predictions based on these learned patterns.  

Trained on almost 30 years of ORAS6 reanalysis data, ML-Ocean predicts the evolution of temperature, salinity and currents throughout the ocean depth (more than 6km), along with variations in sea-surface height and key sea-ice variables. 

The model uses a graph transformer architecture to predict daily average values for a variety of lead times, with the current focus on medium-range timescales up to 15 days. Trained using atmospheric forcing from the ERA reanalysis, ML-Ocean can be run in different configurations: coupled to an atmospheric model such as the AIFS, or forced by atmospheric data from reanalysis, physics-based models, or machine-learning-based systems.  

How well does it work? 

ML-Ocean is compared to the physics-based ocean model NEMO, which is run operationally as part of a coupled prediction system at ECMWF. Here we run both models under ERA5 forcing, giving them the best available atmospheric conditions.  

We see that the models give comparable skill. For temperature, which is key for skilful coupled predictions, ML-Ocean outperforms NEMO, especially at longer lead times. For velocity, skill is comparable to NEMO, particularly towards the middle of the forecast period. 

Comparisons with profile observations show that ML-Ocean produces lower errors than NEMO throughout depth. By training the model on reanalysis data, we achieve improved skill compared to physics-based models, as well as a much faster and more efficient model. 

What’s next? 

The ocean plays a key role in weather and climate predictions across all timescales. Here we’ve shown how skilful the model is over the medium range, and further work will focus on longer timescales, from weeks to months ahead. Predicting over these timescales requires a move towards probabilistic predictions, where the model generates multiple predictions, to provide not only the most likely scenario, but a full spectrum of potential scenarios, and their likelihoods. This work is part of the efforts to produce a data-driven Earth system model with in the DestinE initiative, by coupling various earth system components including ML-Ocean.