| Popis: |
The parameterizations of submesoscale ($<10$km) ocean surface flows are critical in capturing the subgrid effects of vertical fluxes in the ocean mixed layer, yet they struggle to infer the full-complexity of these fluxes in relation to the large scale variables that help set them. In this work, we present a data-driven approach for the submesoscale parameterization, utilizing information from the high-resolution submesoscale-permitting MITgcm-LLC4320 simulation (LLC4320). The new parameterization is given by a Convolutional Neural Network (CNN) trained to infer the subgrid mixed layer vertical buoyancy fluxes as a function of relevant large scale variables. In contrast to previous physics-based approaches, such as the Mixed Layer Eddy (MLE) parameterization, here the CNN infers vertical fluxes that are directly computed from the LLC4320 data, where the submesoscales are resolved down to a resolution of approximately 2km. The CNN has significantly high skill compared with the MLE parameterization, which we demonstrate over a wide range of dynamical regimes and resolutions. We find that the improved skill can be attributed to learned physical relationships between submesoscale fluxes and the large scale strain field, currently missing from submesoscale parameterizations in General Circulation Models. |
