Seasonal Arctic sea ice forecasting with probabilistic deep learning

Area of Science:

• Climate Science
• Artificial Intelligence

Background:

• Anthropogenic warming is causing significant year-round reductions in Arctic sea ice extent.
• This sea ice loss has profound impacts on Arctic communities, ecosystems, and global climate patterns.
• Accurate seasonal sea ice forecasts are crucial for mitigation and adaptation strategies.

Purpose of the Study:

• To develop and evaluate a novel deep learning system for probabilistic, seasonal sea ice concentration forecasting.
• To assess the performance of this system against existing dynamical models at extended lead times.

Main Methods:

• A probabilistic deep learning model, IceNet, was developed.
• IceNet was trained using climate simulations and observational sea ice data.
• The system forecasts monthly-averaged sea ice concentration maps for the next six months.

Main Results:

• IceNet demonstrates improved accuracy in seasonal sea ice forecasts compared to state-of-the-art dynamical models.
• The system shows particular strength in predicting summer sea ice conditions, including extreme events.
• The forecasting range of accurate sea ice prediction is advanced by IceNet.

Conclusions:

• IceNet represents a significant advancement in sea ice forecasting capabilities.
• This deep learning approach enhances our ability to predict critical changes in Arctic sea ice.
• Improved seasonal forecasts can support the development of tools to manage risks from rapid Arctic sea ice loss.