Inferring mechanisms of ocean circulation change in CMIP models using deep learning

Abstract:

A neural network is presented that offers unprecedented insight into ocean model dynamics using only data from the surface available in the CMIP models. The ocean is a key component in climate, transporting and storing heat and tracers like carbon, but enough data to understand the dynamical causes of change are lacking and impractical to store in the CMIP context. The neural network uses readily available climate model data from the surface wind stress, sea level and ocean topography, and trains on a global geography of dynamical classes identified using unsupervised learning in Sonnewald et al. (2019), and is validated independently. The network returns a global geography from a given CMIP model of readily interpretable dynamical classes within the barotropic vorticity framework. CMIP models ESM and CM3 are used to demonstrate changes in historical and 4xCO2 scenarios. For the 'historical' scenario, the predicted geographies are similar to those from the two models with fully known dynamics. The 4xCO2 scenario shows changes in the Southern Ocean and in the wind driven gyres, suggesting that the depth coherence of the flow has changed and impacted heat storage. Differences in the geographical extent of the classes related to up or down welling are also observed, with implications for the global overturning.