Trained on six major climate simulation datasets, a convolutional neural network identifies states of the Arctic Ocean that are conducive to a sudden loss of sea ice within a few years.
In climate simulations subject to an increase in greenhouse gases, the Arctic can lose a significant part of its September sea ice in just a few years. This type of episode has not yet been clearly observed in the satellite survey, which is too short. A study published on arXiv on October 12, 2026 trains a convolutional neural network on six large sets of climate models to diagnose vulnerability to these rapid losses.
Thin ice and warm ocean
Tested in cross-validation on models excluded from training, the network correctly separates high-risk states from those at low risk. The indicators it favours are permanently thin ice, the remaining extent of sea ice and high heat stored in the upper layers of the ocean.
Applied to observations of the satellite era, it signals an increased risk just before the stall of about 2007, then an increasing risk from about 2015 until today. The authors point out that such an episode would have serious consequences for Arctic ecosystems, coastal communities and global shipping.
Read the full article on the original source
Go to the source →