Oceananigans adjoint model, tangent linear model, and variational data assimilation?

[ThomasHaine]

Hi Folks,

Does anyone have experience with, or ideas about, building adjoint models and tangent linear models for Oceananigans? I'm thinking about using Oceananigans for sensitivity analysis and 4D variational data assimilation. Presumably, autodiff tools in Julia, like JuliaDiff, should make this relatively easy. What are the prospects for getting this to work?

I'd be happy to help, e.g., create a tutorial example (although I'm no Julia expert).

Thanks in advance!

[glwagner]

@ThomasHaine there's a short and long answer to this question. In short, we are working on automatic differentiation using Enzyme and Reactant via the DJ4Earth project, and more information can be found here: https://github.com/DJ4Earth and here: https://dj4earth.github.io/. We may have more to share about this soon.

[ThomasHaine]

Great! Thanks @glwagner. I'm interested in staying informed.

[glwagner]

@ThomasHaine there's a preprint out that demonstrates differentiability of a channel setup with bathymetry: https://d197for5662m48.cloudfront.net/documents/publicationstatus/290740/preprint_pdf/03e822b41dec8485587db231a3a9d3bf.pdf

I'm working on a simple single column example in #5013 .

cc @jlk9 @giordano @wsmoses @heimbach