Sea Ice Webinar Series
CMCC-Bo meeting room (2nd floor) and via Zoom
April 12, 2024 at 11:00 CEST

Title: Towards a machine-learned sea ice model parameterization from data assimilation increments

Abstract:
Errors in sub-grid scale climate model parameterizations often lead to biases in numerical forecasts, and increased uncertainty in future projections. In recent years there has been a proliferation of studies investigating whether Machine Learning (ML) can help to improve the representation of these sub-grid processes, and hence improve prediction capabilities within models. In this presentation, I outline how Data Assimilation (DA) is a natural framework for extracting the systematic component of model errors, through the ‘analysis increments’, and subsequently outline how we can use ML techniques to predict these errors. In other words, how we can use ML to emulate a DA process which is independent of observations. I subsequently use the large-scale sea ice model, SIS2, to show how an ML model trained to predict sea ice DA increments can systematically reduce global sea ice biases over a 5-year simulation period. Furthermore, I show how DA lends to an efficient framework for augmenting training data for ML models, leading to improved online performance. These results suggest that this DA+ML workflow has the potential to reduce systematic biases in operational forecast settings, such as seasonal-to-decadal prediction, and also provide new insights into sub-grid scale model parameterizations.

Speaker: William Gregory (Princeton – GFDL)