The use of neural networks to emulate the dynamics of the Earth system (ocean, atmosphere) is of critical importance since it enables efficient and operational forecast, data assimilation, and affordable uncertainty quantification. This field has been very productive in numerical weather forecasts but relatively fewer results were demonstrated in the oceans. This paper could be a reference for mesoscale ocean dynamics emulation.

This paper is of high quality, it exemplifies the complexity of detecting ocean eddies (eulerian versus lagrangian, importance of the code parameters of the regions considered), it acknowledges the fact that not all ocean eddies are the same (dynamically talking), and that their effect on plankton is complex. The method presented is well-described, powerful and would allow digging into these biophysical interactions in many different oceanographic settings. It shows that the common view following seminal papers by e.g. McGillicuddy and D. Chelton that consists of "cyclonic eddies -> isopycnals shaoling -> increase phytoplankton biomass" versus "anticyclonic eddies -> isopycnal deepening -> decreased phytoplankton biomass" is too simplistic and needed to be updated.