Atmosphere Ocean Science Colloquium
Using deep learning to predict systematic model error from sea ice data assimilation increments in a fully coupled climate model
Speaker: Will Gregory, GFDL/Princeton
Location: Warren Weaver Hall 1302
Date: Wednesday, November 9, 2022, 3:30 p.m.
Synopsis:
The so-called “analysis increments” from Data Assimilation (DA) provide a unique insight into the systematic errors of a climate model; errors which can be attributed to either missing physics within the model, erroneous parameterisations of pre-existing physics, or some combination of the two. Here, we hypothesise that such systematic errors are state-dependent, which provides a framework to learn functional mappings from state vector to analysis increment (and hence model error), and subsequently promotes the feasibility of data-driven model parameterisations.
The Seamless system for Prediction and EArth system Research (SPEAR) is the Geophysical Fluid Dynamics Laboratory’s (GFDL's) latest-generation climate model for seasonal-to-decadal prediction and projection, and currently has an experimental sea ice data assimilation system which assimilates and updates ice concentrations at 5-day intervals. In this work, we use Convolutional Neural Networks (CNNs) to learn the relationship between various model states in the forecast window prior to DA (days 1 through 4), and the sea ice concentration increments from DA on day 5. We find that the CNN is able to make accurate predictions of the DA increments using various inputs including sea ice concentration, thickness, and velocities; with particularly high skill occurring in areas which are dominated by biases relating to the sea ice edge location. The sea to the east of Greenland for example shows a root mean squared validation error reduction by as much as 50% in winter with the CNN model. These findings suggest that a machine-learned parameterisation based on sea ice DA increments has the potential to substantially reduce sea ice biases in the free-running SPEAR model.