Leveraging Machine Learning algorithms for improved Arctic sea-ice prediction using the Met Office suite of models

Polar remote sensing of the sea-ice covered oceans was transformed when researchers at the Centre for Polar Observation and Modelling (CPOM) at UCL first realised that it was possible with satellite altimetry to measure the elevation difference – or freeboard – between the sea ice and the level water to deduce the total ice thickness as well as the ocean circulation patterns in the polar oceans. Using this technique (1) it has been possible to produce maps and along-track products of the sea ice thickness (SIT) from several radar altimeters. This also enabled the data assimilation (DA) of these satellite-derived SIT observations in the Met Office’s coupled ocean-sea ice forecasting system (2, 3), as part of a previous CASE studentship between the Met Office and UCL.

More recently, the combination of deep learning tools with the physical understanding of the sea ice-atmosphere interactions made it possible to extend the sea ice thickness observation products to summer months (4, 5). These new products have revolutionised our ability to initialise climate models and proved that early summer sea ice thickness maps have the potential to break the ‘Spring Sea ice predictability barrier’ – a term coined to indicate the challenge in forecasting Autumn sea ice behaviour before May. Since then, Machine Learning (ML) techniques have become ubiquitous in Arctic remote sensing, climate modelling and sea-ice forecasting applications.

In this CASE project, the student will explore the use of state-of-the-art ML algorithms developed in collaboration between the UCL, Met Office and the Alan Turing Institute. The specific objectives of the project are listed below:

1. The development of a long ocean-sea ice reanalysis run, using the latest Met Office sea-ice DA capabilities (2, 3) and the latest sea-ice model component (6), to produce a consistent training dataset for the deep learning techniques. For example, the SIT DA will also include the recently developed summer observation products, providing better SIT initial conditions all year round. There are also plans at the Met Office to develop a regional, high-resolution Arctic model, which can also be used as a ML training dataset for this project.

2. The implementation of deep learning algorithms to improve Met Office’s short range and seasonal forecasts of different sea ice variables, including sea ice concentration and thickness.

3. Comparison of the ML-driven sea-ice forecasts with the dynamical sea-ice model forecasts from the Met Office, using a range of forecast verification metrics and looking at different Arctic regions, particularly those near the ice edge.