Stochastic-Deep Learning Parameterization of Subgrid Ocean Processes in the MOM6 Ocean Model
Abstract
We implement and evaluate a stochastic-deep learning model as a parameterization in a numerical ocean circulation model and discuss the transferability, robustness, and efficiency of the network. Either missing physics, poor subgrid parameterizations of unresolved physical processes, or discretization errors can introduce model biases and errors. In this work, we focus on reducing model biases by parameterizing unresolved and partially-resolved mesoscale ocean eddies. We implement a convolutional neural network model that was developed and trained to represent the statistics of mesoscale eddies from a global eddy-permitting model (Guillaumin and Zanna, 2021). The training minimizes a negative log-likelihood loss function and models the mean and standard deviation of a Gaussian probability distribution of the subgrid momentum forcing. To examine the transferability, robustness, and efficiency of the model, we construct a system consisting of a conventional ocean circulation model (MOM6) configured to simulate idealized wind-driven double gyres, and the trained data-driven model implemented in Python, and test it against fine-resolution simulations. We investigate the problem from three perspectives. We first evaluate how well the machine learning model trained using sea surface velocity can predict the subgrid forcing at depth. We find performance is not optimal, but scaling of the forcing from the CNN model can improve certain metrics. Second, we assess if the forcing near water-land boundaries is predicted as well as it is in open ocean, and find a dependence on the limited geographic regions used during training. Third, we evaluate the cost of the CNN model and compare it to the cost of MOM6 simulations, to estimate if the cost of our machine learned parameterization is prohibitive for use in realistic global simulations. We go over these topics in depth, discuss the possible limitations, and present some solutions to improve the performance of the CNN model in a comprehensive ocean model.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFMNG16A..05Z