Abstract:Results of coarse-resolution climate models are sensitive to the specification of ocean eddy mixing coefficients. Therefore, it is important to estimate, rationalize and predict eddy diffusivities. Here, we estimate the seasonal variability of surface eddy diffusivities in the Kuroshio Extension region using numerical particles advected by a submesoscale-permitting model solution. We find that both the spatial structure and the domain-averaged value of the particle-based eddy diffusivities have a significant seasonal cycle. We also assess the predictability of cross-stream mixing lengths in this region using the methods of machine learning, suppressed mixing length theory (SMLT), and multiple linear regression (LR). The predictors we choose are all variables from SMLT that represent eddy- and mean-flow properties, and these predictors correlate well with the particle-based cross-stream mixing lengths. We demonstrate that, compared to SMLT and LR, machine learning methods, in particular the random forest (RF) and convolutional neural network (CNN), can better represent both the spatial structure and the domain-averaged value of cross-stream mixing lengths. The skill in predicting the mixing lengths with CNN has much less seasonal variability than that with RF. Our results indicate that the machine learning approach may be useful in future development of eddy parameterization schemes. Plain Language Summary Oceanic eddies influence ocean circulation by stirring and mixing heat and salt, and thus affect climate variability. In coarse-resolution climate models, accurate simulation requires a good parameterization of the eddy mixing process. Estimating, understanding, and predicting the rate of mixing induced by eddies is a vital step toward making such a parameterization. In this study, we estimate and predict the seasonally varying surface eddy mixing rates in the Kuroshio Extension region using a high-resolution ocean simulation, dynamical theory, and machine learning methods. We find that surface eddy mixing rates in this region have significant seasonal variability and we demonstrate that the machine learning method can better represent the eddy mixing process than the conventional dynamical theory. Thus, our results indicate that for parameterizing eddy mixing, a machine learning approach may be useful and that one should use seasonally varying eddy mixing coefficients.

Seasonal Surface Eddy Mixing in the Kuroshio Extension: Estimation and Machine Learning Prediction