AI enables prediction of El Niño events in South Atlantic Ocean months in advance

April 14, 2025 This article has been reviewed according to Science X's editorial process and policies . Editors have highlightedthe following attributes while ensuring the content's credibility: fact-checked peer-reviewed publication proofread by Tori Pedersen, Bjerknes Centre for Climate Research The El Niño phenomenon in the South Atlantic and Benguela current, which flows along the west coast of southern Africa, have a significant impact on the tropical Atlantic region, leading to extensive effects on local marine ecosystems, African climates, and the El Niño Southern Oscillation. No one has been able to predict warm events in this region until now.

"We are extremely excited because it was the first time we could actually produce some predictions useful for communities building traditional models and achieving that goal that we had two years before," says Marie-Lou Bachèlery about her paper just published in Science Advances . The tropical Atlantic Ocean is situated between the Brazilian coast to the west and the West African coast to the east. The Central Atlantic Niño is characterized by warm sea surface temperatures concentrated in the central equatorial Atlantic, while the eastern Atlantic Niño features warming in the eastern equatorial Atlantic, near the West African coast.

As a significant component of climate systems, variations in the ocean influence local weather patterns. These changes also impact marine ecosystems and the livelihoods of people who depend on the ocean's fish resources. In a deadlock The South Atlantic is one of the regions which sticks out as a region that has quite strong warming in the ocean.

And this is problematic for a lot of things, like fisheries. This makes it even more important to predict extreme events , like Atlantic and Benguela Niño events. "My idea was to make predictions of those events using climate models .

.. after a year and a half of work, we realized that it did not work and that we were kind of in a deadlock situation," Bachèlery says.

At that time Bachèlery was working at the Geophysical Institute at the University of Bergen. Now she is working at the Euro-Mediterranean Center on Climate Change in Italy. New approach Climate models often struggle to predict warm events in the tropical Atlantic due to their low resolution.

These models fail to accurately represent upwelling dynamics, where wind-driven processes bring deeper, cooler water to the surface. This upwelling requires high-resolution modeling to capture the fine-scale processes involved. The inability to resolve these dynamics leads to significant temperature biases in the region.

These biases create a cascade of errors, affecting surface temperatures, atmospheric coupling, and teleconnections, ultimately resulting in inaccurate predictions of warm events. "With innovative techniques such as machine learning , I started to think of the possibilities and that I knew the region very well. I knew exactly what I needed to put in to predict those events," Bachèlery explains.

Professor Noel Keenlyside was Bachèlery's supervisor and has worked with prediction for many years. The first time he worked on predicting in the Atlantic region was about 15 to 20 years ago. "For the first time, it is actually possible to predict these events and overcome the issue of model errors by using a different approach.

Many people have been trying to predict that area for a few decades. That's why Marie Lou's results are exciting," he says. The study was also carried out under the TRIATLAS-project, which had the goal of assessing the status of the South and tropical Atlantic marine ecosystem and developing a framework for predicting its future changes, from months to decades.

To be able to predict warm events will be very useful for fisheries. "When extreme events occur, managers may limit the fishing in this region, to reduce effects of the additional pressure from the environment," Keenlyside says. When they first got the results, Bachèlery did not believe what she saw.

"I rechecked what felt like a billion times to make sure that we were not over predicting, because that is a common feature in machine learning." To make a prediction, they gave the machine maps of the temperature of the region they were interested in, and then the machine learning model identifies patterns and regions that provide information that led to accurate predictions for the next two months. By accessing this information, they can represent physical mechanisms, as the prediction model is trained on real data.

"The machinery was not doing random stuff; it was actually relying on real physical mechanisms that exist. So that's that was the most interesting part for me," Bachèlery says. The feedback on the results has been extremely positive.

Bachèlery explains that she has gotten comments on the fact that they managed to predict this system, and people ask questions, like what type of algorithm they used. Next steps "Even if the system is not necessarily for their region, the whole technique can be repeated for any other system. And I think people are really excited about that.

" They are now working to make these forecasts available through a dashboard. "We are in dialogue with forecast users from National Institute for Fisheries in Angola (INIP) to further improve and refine the information to their needs. It is particularly pleasing to see that basic research is becoming societally relevant," Keenlyside says.

More information: Marie-Lou Bachèlery et al, Predicting Atlantic and Benguela Niño events with deep learning, Science Advances (2025). DOI: 10.1126/sciadv.

ads5185 Journal information: Science Advances Provided byBjerknes Centre for Climate Research.