Unsupervised Learning Reveals Geography of Global Ocean Dynamical Regions

Dynamically consistent global ocean regions are identified in a simplified barotropic vorticity framework from a twenty-year mean of the Estimating the Circulation and Climate of the Ocean (ECCO) state estimate at 1^o, highlighting regions where surface stress is dominant. Closure of the barotropic vorticity budget provides means of classification and the dynamical interpretation, highlighting where linear theory is appropriate and where non-linear terms are significant. An unsupervised learning algorithm, K-means, is the objective clustering analysis demonstrating five unambiguous regimes with varying term balances and terms of varying sign. The most dominant Cluster 1 covers 57± 2% of the ocean area. Surface and bottom stress terms are balanced there by the bottom pressure torque and the non-linear terms. Cluster 2 covers 18± 0.7%, and is where the beta effect balances the bottom pressure torque. Cluster 3 covers11±0.5%, characterized by a ``Quasi-Sverdrupian'' regime where the beta effect is balanced by the wind and bottom stress term. The small region of Cluster 4 has baroclinic dynamics covering about 5.0±1.9% of the area. Cluster 5 occurs primarily in the Southern Ocean. Residual ``dominantly non-linear'' regions are found areas of rough topography in the Southern Ocean and along western boundaries. The dominantly non-linear areas are the subject of future work applying the unsupervised learning at higher resolution to assess the integrated effect of the smaller scales.