Linear Inverse Modeling for Arctic Sea Ice Prediction

Abstract:

Global climate models (GCM) are powerful tools for probing dynamical drivers of Arctic sea-ice variability. However, GCMs are computationally expensive making them impractical for running long simulations and large ensembles. In order to exploit the dynamical information contained in GCMs in a computationally efficient way, we explore the use of Linear Inverse Modeling (LIM) for Arctic sea-ice prediction. LIMs decompose a multivariate dynamical system into the sum of the linearized dynamics and white noise, which model the slow-varying linearly predictable and high frequency unpredictable processes respectively. The result is a linear model that emulates the dynamics of a subspace of a GCM that can produce large ensemble simulations of the multivariate climate system. We focus on LIM skill in predicting Arctic sea-ice concentration and thickness on monthly to seasonal timescales. We first show that a LIM trained on GCM last millennium simulations is skillful in predicting preindustrial Arctic sea-ice conditions for both in and out-of-sample data. We explore how this performance compares to damped persistence and how the LIM handles the known asymmetries in the seasonality of Arctic sea ice. We then use the LIM to test hypotheses regarding dynamical conditions driving the early 20th century warming (ETCW) in the Arctic. Particularly, we investigate the role of the Pacific versus Atlantic basins as well as ocean heat transport and ice albedo feedback during the ETCW. Moreover, we use the LIM to explore optimal drivers of sea-ice variability.