The sea ice in the Arctic has several key functions in the Earth's climate system: it reflects large parts of the solar radiation, thus slowing down global warming and it drives global ocean and air currents, which has a direct impact on the weather outside our window. How the sea ice retreats or expands depends largely on the energy balance of the Arctic Ocean surface. However, there is a lack of data from direct observations, meaning that studies have to rely on modelling with reanalysis data from the past in order to make predictions for the future. However, these can be distorted by the data gap. Researchers at the Alfred Wegener Institute have now developed an AI-supported model that can correct these biases and thus contribute to a better understanding of Arctic climate processes. They present their method in the scientific journal Geophysical Research Letters.
The new method is based on machine learning so that the model can independently recognize patterns from a data set and make predictions. “We trained a neural network with observational data from several large Arctic field campaigns, for example from the MOSAiC expedition,“ says lead author Dr Akil Hossain from the Alfred Wegener Institute (AWI). The climate modeller and his colleagues applied the network, called SEBai (Surface Energy Budget ai), to the ERA5 climate model, which is frequently used to reconstruct energy and heat fluxes at the surface of the Arctic Ocean.
What they found was that SEBai can significantly improve the predictions of the most important surface fluxes and the 2-metre temperature. “We were able to reduce the error rate for the sunlight absorbed at the surface by 40 per cent. For the total energy budget of the surface by more than half,“ says Akil Hossain. Specifically, the results show that winter temperatures in the Arctic are actually around 4 degrees colder than the ERA5 dataset suggests, while summer temperatures are around 1 degree colder. This indicates that the Arctic Ocean actually cools down less in winter and warms less in summer, which may make the Arctic sea ice more vulnerable to climate change than previously assumed.
“Our model shows that machine learning combined with sparse observational data can significantly improve estimates of Arctic surface fluxes and near-surface temperature,“ summarizes Akil Hossain. “This allows us to make sea ice and ocean simulations more precise and to better understand the climate processes in the Arctic that are so important to us.“
Original publication
Akil Hossain, Paul Keil, Harsh Grover, et al. Machine Learning Eliminates Reanalysis Warm Bias and Reveals Weaker Winter Surface Cooling over Arctic Sea Ice. Geophysical Research Letters, 53, e2025GL121289. 19 June 2026. DOI: https://doi.org/10.1029/2025GL121289
