For more than 20 years, around 4,000 autonomous profiling floats have been drifting through the ocean. They form part of the international Argo program (Argo – global array of profiling floats). At regular intervals, they descend to a depth of 2,000 meters and, as they rise, measure parameters such as temperature, salinity, and pressure. Once they have reached the sea surface again, they transmit this data via satellite to the Argo network, which is available to researchers worldwide. Hardly any other observation system has transformed ocean research so profoundly in recent decades.
Now, a study shows that this data can be analyzed far beyond its original intended use. “We wanted to know whether we could use the scattered Argo measurements to gain insights into large-scale circulation systems that, until now, could only be recorded through very extensive measurement campaigns,” says Yannick Wölker, until recently a Ph.D. student in the Ocean Dynamics research unit at GEOMAR and lead author of the study. “Artificial intelligence opens up new possibilities here. Combining machine learning with established physical models allows us to get more out of existing measurement data and better understand how key circulation systems work.”
The study focuses on the Atlantic Meridional Overturning Circulation, or AMOC for short. This vast circulation system acts like a conveyor belt, transporting warm surface water northwards, where it cools, sinks into the depths and flows back south as cold deep water. The AMOC transports large amounts of heat and influences weather and climate, including in Europe. How stable this system is and how it is changing under the influence of climate change is one of the key questions currently discussed in climate research. So far, however, direct observations have relied on just a few fixed measurement series in the Atlantic, which are technically challenging and expensive to carry out.
To tackle this problem, the researchers combined two approaches: observational data and model calculations. They trained a machine learning algorithm (a subfield of artificial intelligence in which computers recognize patterns in data) using high-resolution ocean simulations. In the process, the model learnt how typical ocean current patterns correlate with temperature and salinity profiles. They then applied this knowledge to real Argo data from the Atlantic. This approach makes it possible to infer large-scale current strengths from point measurements—in particular the so-called geostrophic component of the circulation (determined by the distribution of temperature and salinity), which has hitherto been difficult to measure continuously.
The results show that the estimates calculated from the Argo data agree well with established observational series and model simulations. The study thus provides an important methodological building block for ocean observations.
At the same time, the authors emphasize the limitations of their approach. The method is based on model assumptions and only covers specific time windows. Very short-term fluctuations, as well as the long-term decline in the AMOC, can only be determined to a limited extent. The new methodology is therefore not intended to replace existing measurement systems, but rather to complement and optimize them effectively.
“Our approach is no substitute for direct measurements in the ocean,” says Yannick Wölker. “But it can help to make better use of existing data and bridge gaps in observations.”
The method could play an important role in planning future observation networks. It demonstrates how global programs such as Argo can gain greater significance through modern data analysis—and how additional infrastructure in the ocean can be strategically deployed.
“Particularly with regard to long-term climate monitoring, we need to consider how to design observations that are efficient, robust and internationally coordinated,” says Prof. Dr. Arne Biastoch, a professor at GEOMAR and co-author of the study. “Artificial intelligence methods can help to ensure that strategic decisions regarding future ocean measurements are made in the best possible way.”
The work was carried out in collaboration as part of the Helmholtz School for Marine Data Science (MarDATA), a Ph.D. program involving researchers in Kiel and Bremen, in which Ph.D. students are supervised by marine scientists and computer scientists jointly.
Original Publication:
Wölker, Y., W. Rath, M. Renz, and A. Biastoch, 2025: Estimating the AMOC from Argo Profiles with Machine Learning Trained on Ocean Simulations, Ocean Sci., 21, 3541–3562, doi.org/10.5194/os-21-3541-2025.
