Early Warning of Ocean Tipping Points: A Deep Learning model approach

In the study of nonlinear dynamics, tipping points have been a major focus of research. They describe a threshold where gradual changes in external forcing, e.g. increasing CO2 emissions can lead to an abrupt and persistent transition. This is a concern in ocean sciences due to the strong coupling of the ocean and climate. The capability to provide an early warning of a tipping element is desirable, providing time to mitigate and adapt.

Specific regions of the ocean are of more concern for potential tipping elements than others, a key region being the sub polar gyre. This is a basin-scale cyclonic gyre in the North Atlantic, driven by wind and buoyancy forcing. It is crucial in the formation of North Atlantic Deep Water, a main contributor to the lower branch of the Atlantic Meridional Overturning Circulation (AMOC). If this deep convection process collapses, then cascading changes in sea ice, atmospheric circulation, ocean circulation and sea level, and the terrestrial ecosystem are expected.

Machine Learning approaches suggest that a generalised deep learning (DL) model could potentially provide a robust and high confidence solution to predicting tipping points. We have applied an existing generalised DL model to a large ensemble of historical and future climate projections (1950-2100) based on the HadGEM3 Atmosphere-Ocean-Sea-Ice-Land model under the SSP370 future shared socioeconomic pathway scenario. Using change point analysis to identify tipping points in this ensemble, the DL was provided with timeseries of several parameters, leading up to but not including the identified tipping points.  We then assessed the ability of the DL to predict the tipping points based on the chosen parameter timeseries across a number of specific geographic regions.

Mixed Layer Depth and Sea Surface Height were the most effective parameters and there was a large variation in the effectiveness across different regions, with some (Labrador Sea) being much better than others (Irminger Sea). While current DL models are not yet capable of robust tipping point detection there is clear promise in continuing to refine this method with new DL models specifically created for ocean surface and subsurface parameters, such as MLD and temperature and salinity depth profiles.