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

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.

Impact of Parameterized Isopycnal Diffusivity on Shelf‐Ocean Exchanges Under Upwelling‐Favorable Winds: Offline Tracer Simulations Augmented by Artificial Neural Network

“…As with previous authors invoking machine learning techniques to parameterize subgrid-scale processes in ocean and climate models (Bolton & Zanna, 2019;Duo et al, 2019;George et al, 2021;Guan et al, 2022;Manucharyan et al, 2021;Partee et al, 2022;Rasp et al, 2018;Salehipour & Peltier, 2019;Yuval & O'Gorman, 2020;Zanna & Bolton, 2020Zhu et al, 2022), we have leveraged the ANN to formulate closure forms of the isopycnal eddy diffusivity (i.e., a diffusivity that can be formulated or inferred solely from the mean flow and topographic information) based on two approaches. The first approach is to couple the ANN-learned EKE with the physics-based, bathymetry-aware scaling for κ Redi .…”

“…For completeness, an additional ANN is trained to directly infer the cross‐slope isopycnal eddy diffusivity from the mean flow and topographic quantities, leading to a purely data‐driven eddy closure that enables the parameterized model to nearly reproduce the tracer solutions of the eddy‐resolving model. This work is intended as a contribution to the growing body of literature devoted to parameterizing the fine‐scale processes in predictive ocean and atmospheric models using machine learning approaches (Bolton & Zanna, 2019; George et al., 2021; Guan et al., 2022; Manucharyan et al., 2021; Partee et al., 2022; Rasp et al., 2018; Salehipour & Peltier, 2019; Yuval & O’Gorman, 2020; Zanna & Bolton, 2020, 2021; Zhu et al., 2022).…”

Nonlocality of scale-dependent eddy mixing at the Kuroshio Extension

“…For example, the suppressed mixing length theory from Ferrari and Nikurashin (2010) expresses diffusivity as a function of local mean flow and eddy properties. However, recent work shows that the value of eddy diffusivity depends on both local and nonlocal flow fields (Chen and Waterman, 2017;Guan, 2022), indicating the need of developing nonlocal eddy parameterization schemes. In fact, subgrid parameterization schemes for several other physical processes (e.g., diapycnal mixing and atmospheric boundary layer) have already included the nonlocality effect (Large et al, 1994;Frech and Mahrt, 1995;Brown and Grant, 1997;Noh et al, 2003;Hong et al, 2006;Inoue et al, 2010;Chen et al, 2021).…”

“…Recently, several studies show that the nonlocality of total eddy mixing is non-negligible in idealized western boundary extensions or at the KE region (e.g., Chen et al, 2014;Chen and Waterman, 2017;Guan et al, 2022). For example, using a barotropic quasigeostrophic model and Lagrangian particles, Chen and Waterman (2017) estimated the nonlocality for total mixing in an idealized western boundary current jet.…”

“…They demonstrated that the nonlocality for total mixing is prevalent in the domain, and it is stronger within the jet compared to the jet flanks. Guan et al (2022) estimated the nonlocality for total mixing in the KE region using a high-resolution simulation, MITgcm llc4320. They found that the domain-averaged degree of mixing nonlocality is larger than 200 km.…”