The SuperIce Webinars are a series of scientific online presentation and discussion on the topic of sea ice and machine learning. They are organized by the SuperIce project, which is led by the Nansen Center. The webinars are open to everyone. For this first webinar, we will have two presentations.
William Gregory – From component to coupled: evaluating the performance of a machine-learned sea ice bias correction scheme in fully-coupled seasonal predictions
Recent work has shown that Machine Learning (ML) algorithms trained to predict sea ice Data Assimilation (DA) increments can systematically reduce sea ice biases when used as an online bias correction scheme. These findings motivate two questions: 1) Can this ML model generalise to higher implementation frequencies than the original DA experiment? 2) Can this ML model generalise to a different configuration of the numerical model? In this presentation, I discuss recent challenges and successes in addressing these questions. Starting with 5-day implementation frequency in forced ice-ocean simulations, I show that (with some modifications to the network inputs) the ML corrections can generalise to each timestep of the numerical model (30 minutes). At this higher implementation frequency, the online performance is generally improved and imprints of model shock, which were present in the 5-day implementation, are removed. I then show results of using the ML model to bias correct July 1st initialised seasonal sea ice forecasts in the fully-coupled GFDL SPEAR model. Prediction skill is improved between July and September in the Arctic, although the network results in delayed September freeze-up, leading to increased biases at the start of the Arctic growth season. In the Antarctic, the network may be triggering ice-albedo feedbacks, leading to rapid ice melt between October and January. Network ablation studies reveal that this may be due to specific network inputs which are out of sample in the fully-coupled model. These findings suggest that, with some refinement to the network, the ML-based correction scheme has the potential to improve seasonal sea ice predictions in both hemispheres.
Nicholas Williams – Super-resolution of satellite observations of sea ice thickness using diffusion models and physical modeling
We will present new developments of the Norwegian Climate Prediction Model (NorCPM) to improve reanalysis and prediction of Arctic sea ice. NorCPM is developed at the Bjerknes Centre for Climate Research in Norway and a combination of the Norwegian Earth System Model (NorESM) and the Ensemble Kalman Filter (EnKF) for the purposes of reanalysis and prediction. The assimilation of sea ice thickness is newly implemented using observations from ESA CCI (ENVISAT) and C2SMOS. This is the first time that the ESA CCI dataset has been assimilated in a sea ice model, and it shows that it can be useful for the 2002-2010 period where we currently lack SIT observations. For reanalyses we find that the assimilation can constrain the error substantially to under 1 m for the ESA CCI period and under 0.4 m under the C2SMOS period and removing the strong bias for thick ice in NorCPM. For prediction we find that the pan-Arctic sea ice extent in summer is particularly improved, especially for the minimum sea ice extent in September. There are also regional improvements within many of the Arctic marginal seas, especially in the Central Arctic and Beaufort Sea, where RMSE is reduced and correlation to observations in summer is increased.
