Work

An online-learned eddy parameterization in a differentiable shallow water model

Year: 2026
Type: preprint
Abstract: Ocean general circulation models rely on parameterizations to approximate subgrid-scale processes. While parameterizations are traditionally physics-based, machine learning has recently provided an alternative, data-driven approach for developing parameterizations. However, most machine learning-based parameterizations are learned offline due to the lack of differentiable general circulation models. Here, we train an eddy momentum parameterization represented by a neural network online in an idealized barotropic gyre model. We develop the model as fully differentiable in Julia with the automatic differentiation tool Enzyme. The network weights are trained using a loss function that has the same form as in adjoint-based data assimilation, with synthetic observations generated by a fine-resolution simulation. A range of diagnostics evaluate the performance of the parameterized coarse-resolution model. We find improved energetics, physically realistic dynamics, and numerical stability for multi-decadal roll-outs. Importantly, the online-learned parameterization absorbs artifacts from the numerical discretization, such as the time-stepping scheme used, which is invisible to offline-learned parameterizations. We suggest that this is key to improved numerical stability of machine-learned parameterizations. Our research suggests that online-trained neural network parameterizations can be an effective solution to the subgrid-scale parameterization and model calibration problem.
Language: English
Subfield: Oceanography
Open Access status: green
OpenAlex ID: W7225894586
Cites:
Cited by:
Related to: 10