Deep Learning of Unresolved Turbulent Ocean Processes in Climate Models

Current climate models do not resolve many nonlinear turbulent processes, which occur on scales smaller than 100 km, and are key in setting the large-scale ocean circulation and the transport of heat, carbon and oxygen in the ocean. The spatial-resolution of the ocean component of climate models, in the most recent phases of the Coupled Model Intercomparison Project, CMIP5 and CMIP6, ranges from 0.1∘ to 1∘ (Taylor et al. 2012; Eyring et al. 2016b). For example, at such resolution, mesoscale eddies, which have characteristic horizontal scales of 10–100 km, are only partially resolved – or not resolved at all – in most regions of the ocean (Hallberg 2013). While numerical models contribute to our understanding of the future of our climate, they do not fully capture the physical effects of processes such as mesoscale eddies. The lack of a resolved mesoscale eddy field leads to biases in ocean currents (e.g., the Gulf Stream or the Kuroshio Extension), stratification, and ocean heat and carbon uptake (Griffies et al. 2015). To resolve turbulent processes, we can increase the spatial resolution of climate models. However, we are limited by the computational costs of an increase in resolution (Fox-Kemper et al. 2014). We must instead approximate the effects of turbulent processes, which cannot be resolved in climate models. This problem is known as the parameterization (or closure) problem. For the past several decades, parameterizations have conventionally been derived from semi-empirical physical principles, and when implemented in coarse resolution climate models, they can lead to improvements in the mean state of the climate (Danabasoglu et al. 1994). However, these parameterizations remain imperfect and can lead to large biases in ocean currents, ocean heat and carbon uptake. The amount – and availability – of data from observations and high-resolution simulations has been increasing. These data contain spatio-temporal information that can complement or surpass our theoretical understanding of the effects of unresolved (subgrid) processes on the large-scale, such as mesoscale eddies. Efficient and accurate deep learning algorithms can now be used to leverage information within this data, exploiting subtle patterns previously inaccessible to former data-driven techniques. The ability of deep learning to extract complex spatio-temporal patterns can be used to improve the parameterizations of subgrid scale processes, and ultimately improve coarse resolution climate models.