Implementation and Evaluation of a Machine Learned Mesoscale Eddy Parameterization Into a Numerical Ocean Circulation Model

Abstract

We address the question of how to use a machine learned (ML) parameterization in a general circulation model (GCM), and assess its performance both computationally and physically. We take one particular ML parameterization (Guillaumin & Zanna, 2021, https://doi.org/10.1002/essoar.10506419.1) and evaluate the online performance in a different model from which it was previously tested. This parameterization is a deep convolutional network that predicts parameters for a stochastic model of subgrid momentum forcing by mesoscale eddies. We treat the parameterization as we would a conventional parameterization once implemented in the numerical model. This includes trying the parameterization in a different flow regime from that in which it was trained, at different spatial resolutions, and with other differences, all to test generalization. We assess whether tuning is possible, which is a common practice in GCM development. We find the parameterization, without modification or special treatment, to be stable and that the action of the parameterization to be diminishing as spatial resolution is refined. We also find some limitations of the machine learning model in implementation: (a) tuning of the outputs from the parameterization at various depths is necessary; (b) the forcing near boundaries is not predicted as well as in the open ocean; (c) the cost of the parameterization is prohibitively high on central processing units. We discuss these limitations, present some solutions to problems, and conclude that this particular ML parameterization does inject energy, and improve backscatter, as intended but it might need further refinement before we can use it in production mode in contemporary climate models.

Parameterizing Vertical Mixing Coefficients in the Ocean Surface Boundary Layer Using Neural Networks

“…The successful use of small neural networks as efficient surrogate models of SMCs proves that we can replicate the behavior of complex models with high fidelity. Increasing the network size will be explored in the future and will likely require using GPUs for implementation in OGCM (Zhang et al, 2023). 3.…”

“…The successful use of small neural networks as efficient surrogate models of SMCs proves that we can replicate the behavior of complex models with high fidelity. Increasing the network size will be explored in the future and will likely require using GPUs for implementation in OGCM (Zhang et al., 2023). The performance of the modified vertical mixing scheme in a coupled model (atmosphere‐ocean‐ice) may not show the same impact on model bias as observed in this forced ocean‐ice model.…”

“…Partee et al (2022) trained a deep neural network to learn subgrid kinetic energy of oceanic mesoscale eddies from a high resolution OGCM to improve their representation in a lower resolution OGCM. Convolutional neural networks (CNNs) have been used to predict parameterizations of ocean momentum backscatter in a variety of models (Bolton & Zanna, 2019;Guillaumin & Zanna, 2021;Zanna & Bolton, 2020) and have been implemented in an ocean primitive equation model (Zhang et al, 2023). Gregory et al (2023) recently employed CNNs to learn data assimilation increments for sea-ice and showed that networks could be used to reduce biases in sea-ice.…”

Deep Learning of Systematic Sea Ice Model Errors From Data Assimilation Increments

“…Looking to similar studies, Guillaumin and Zanna (2021) found that implementing a fully CNN with 8 convolutional layers as a stochastic parameterization into an idealized shallow water model resulted in a 25% increase in the run time, compared to an unparameterized simulation. C. Zhang et al (2023) also found that the cost of doing inference with this same network as a parameterization in MOM6 was 10 times more expensive than the CPU cost of the simulation itself. Although we effectively have 8 convolutional layers when considering both networks A and B, we can still expect much lower computational overheads given that ours is a deterministic model (i.e., we predict a single output at each grid point for each ΔSICN, rather than a, potentially larger, number of parameters which describe a distribution of values), and that our kernel size for network B is 1 × 1 in each layer.…”

Machine Learning for Online Sea Ice Bias Correction Within Global Ice‐Ocean Simulations

“…Typically this is achieved by training an ML model to learn a functional mapping which characterizes the impact of subgrid processes on resolved scales, by training on high resolution simulations or observational data. Significant effort is currently being afforded to the development of these ML parameterizations in the context of for example, ocean turbulence, with early results (Frezat et al., 2022; Kurz et al., 2023; Ross et al., 2023; Zanna & Bolton, 2020; C. Zhang et al., 2023) highlighting their potential to improve important climate statistics, such as eddy kinetic energy at large scales, over their traditional physics‐based counterparts. Meanwhile, only recently have studies begun to investigate ML‐based subgrid sea ice parameterizations.…”