[Paper Review] Implementation and Evaluation of a Machine Learned Mesoscale Eddy Parameterization into a Numerical Ocean Circulation Model

Cheng Zhang, Pavel Perezhogin+4 more|arXiv (Cornell University)|Mar 2, 2023

Meteorological Phenomena and Simulations4 citations

TL;DR

This study implements and evaluates a deep convolutional neural network-based machine learning parameterization for mesoscale eddies in the MOM6 ocean circulation model. The parameterization successfully injects energy and improves backscatter, but faces challenges in computational cost, boundary performance, and depth-dependent tuning, indicating promise yet requiring refinement for climate model deployment.

ABSTRACT

We address the question of how to use a machine learned parameterization in a general circulation model, and assess its performance both computationally and physically. We take one particular machine learned parameterization \cite{Guillaumin1&Zanna-JAMES21} 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 general circulation model 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: 1) tuning of the outputs from the parameterization at various depths is necessary; 2) the forcing near boundaries is not predicted as well as in the open ocean; 3) the cost of the parameterization is prohibitively high on CPUs. 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.

Motivation & Objective

• To assess the feasibility and performance of integrating a pre-trained machine learning parameterization into a different ocean circulation model (MOM6) than originally developed for.
• To evaluate the generalization of the ML parameterization across varying spatial resolutions, flow regimes, and model configurations.
• To investigate the practical challenges of implementing ML parameterizations in operational climate models, including tuning, stability, and computational cost.
• To determine whether the ML parameterization can effectively simulate subgrid-scale momentum forcing and improve key physical metrics like energy backscatter.
• To identify limitations in implementation—particularly near boundaries and at different ocean depths—and propose solutions.

Proposed method

• The study implements the stochastic-deep learning parameterization from Guillaumin & Zanna (2021), which uses a convolutional neural network to predict parameters for a stochastic model of subgrid momentum forcing.
• The ML parameterization is embedded online within the MOM6 model as a subgrid-scale forcing term (S_k) in the momentum equation, replacing conventional parameterizations.
• The model is tested across multiple configurations: different spatial resolutions (from 1/10° to 1/40°), distinct flow regimes (e.g., wind-driven double gyre), and with and without tuning of output parameters.
• The parameterization is evaluated using standard oceanographic metrics, including kinetic energy spectra and backscatter, to assess its physical fidelity.
• Computational performance is measured on CPUs and GPUs to assess feasibility for large-scale climate simulations.
• Post-processing and tuning strategies are applied to improve output consistency across depth levels and regions.

Experimental results

Research questions

• RQ1 Can a pre-trained machine learning parameterization for mesoscale eddies be successfully implemented and stabilized in a different ocean model (MOM6) than the one it was trained on?
• RQ2 How does the performance of the ML parameterization vary across different spatial resolutions and flow regimes, indicating its generalization capability?
• RQ3 What are the computational costs of the ML parameterization, and is it feasible for use in production climate models on standard hardware?
• RQ4 To what extent does the ML parameterization improve physical metrics such as energy backscatter and momentum transfer compared to conventional parameterizations?
• RQ5 What are the key implementation challenges—particularly near boundaries and across depth levels—and how can they be mitigated?

Key findings

• The machine learning parameterization is stable when implemented online in MOM6 without modification, demonstrating robustness across different model configurations.
• The parameterization's influence diminishes with increasing spatial resolution, consistent with the expectation that higher resolution resolves more subgrid processes directly.
• Tuning of the ML output at different depth levels is necessary to achieve consistent physical behavior, indicating depth-dependent sensitivity in the model's predictions.
• The parameterization performs poorly near ocean boundaries, where subgrid-scale dynamics are more complex and less represented in the training data.
• The computational cost of the parameterization is prohibitively high on CPUs, making it impractical for large-scale climate simulations without hardware acceleration.
• Despite limitations, the ML parameterization successfully injects energy and improves backscatter, confirming its physical intent is met, though further refinement is needed for operational use.