Predicting the extent of Arctic sea ice in September has significant implications for climate change and shipping in the Arctic. However, seasonal forecasts for September sea ice often encounter a challenge known as the "spring predictability barrier."

To address this issue, a research team led by Prof. LI Xiaofeng from the Institute of Oceanology of the Chinese Academy of Sciences (IOCAS) has developed a new AI model called SICNet_season for seasonal-scale predictions. Their study has been published in Geoscientific Model Development.

Traditional numerical model-based sea ice prediction methods, such as the European Centre for Medium-Range Weather Forecasts (ECMWF)'s SEAS5 and statistical model-based approaches, frequently encounter Arctic sea ice spring predictability barriers.

The SICNet_season model integrates the nonlinear global and local dependencies among sea ice series by Swin-Transformer blocks. This approach allows it to model the long-term relationships between spring sea ice conditions and those in September, leading to more accurate predictions compared to numerical and statistical models.

Experimental results indicate that when using April and May as initial months for predicting September sea ice extent, the SICNet_Season model achieves a 7–10% improvement in predictive skill (ACC) and over 14% better accuracy in sea ice boundary predictions (BACC), significantly reducing the spring predictability barrier. Additionally, spring sea ice thickness is identified as a critical factor, contributing more than 20% to overcoming this barrier.

This study offers an AI-driven solution to the spring predictability barrier in Arctic sea ice forecasts, enhancing the precision of seasonal predictions for September sea ice in the Arctic.

Framework of model SICNet_season. (Image by IOCAS)