Add script for testing FP precision + numerics with ocean climate simulation

[glwagner]

This builds off #72

[NoraLoose]

I'm a bit confused — when testing tracer conservation on the global tripolar grid (even without Reactant and using double precision), I’m seeing a relative error of around 10^{-4}. This is the same order of magnitude we previously observed when using ZCoordinate(). However, with ZStar(), we previously saw very small relative errors of around 10^{-13} when testing with the LatitudeLongitudeGrid.

After running simulations/ocean_climate_simulation.numerics.jl, I double-checked that my test actually uses the z-star vertical coordinate, via:

simulation.model.ocean.model.vertical_coordinate

which returns

Oceananigans.Models.HydrostaticFreeSurfaceModels.ZStar()

Question: Is there something specific about the tripolar grid that prevents ZStar() from working properly?

@glwagner @simone-silvestri

[NoraLoose]

I'm using the free surface defined in src/ocean_climate_model.jl, i.e.,

free_surface = ClimaOcean.OceanSimulations.default_free_surface(grid, fixed_Δt=Δt)

Maybe something is going wrong there?

When I check

simulation.model.ocean.model.free_surface

this returns

SplitExplicitFreeSurface substepping with FixedSubstepNumber(0)

I'm also tagging @navidcy.

[glwagner]

maybe try to use this:

GB-25/simulations/baroclinic_instability_simulation_run.jl

Line 18 in 192b5dc

model = baroclinic_instability_model(arch, resolution=8, Δt=60, Nz=10)

[NoraLoose]

@glwagner you mean we should maybe move away from the data-full ocean climate simulation and back to the baroclinic instability simulation for our numerics tests?

[NoraLoose]

@glwagner you mean we should maybe move away from the data-full ocean climate simulation and back to the baroclinic instability simulation for our numerics tests?

I guess we should continue that work in #16 rather than in this PR.

[simone-silvestri]

Question: Is there something specific about the tripolar grid that prevents ZStar() from working properly?

Not that I know. However, we do not test zstar with a tripolar grid, so something might have broken. Just to know if we broke it recently or never worked, @NoraLoose the tests you did with clima ocean some time ago were with the tripolar grid or the latitude longitude grid?

[NoraLoose]

Not that I know. However, we do not test zstar with a tripolar grid, so something might have broken. Just to know if we broke it recently or never worked, @NoraLoose the tests you did with clima ocean some time ago were with the tripolar grid or the latitude longitude grid?

@simone-silvestri This PR is the first time I’m testing ZStar() with a tripolar grid. Before this, I’ve always used a latitude-longitude grid. No clue if it’s never worked with a tripolar grid or if something broke recently—sorry!

[NoraLoose]

Not that I know. However, we do not test zstar with a tripolar grid, so something might have broken. Just to know if we broke it recently or never worked, @NoraLoose the tests you did with clima ocean some time ago were with the tripolar grid or the latitude longitude grid?

@simone-silvestri, I think the underlying issue is that z-star does currently not work with the SplitExplicitFreeSurface. I'm running these tests locally, and many—possibly all—tests using SplitExplicitFreeSurface, even on RectilinearGrid or LatitudeLongitudeGrid, are failing. (They pass with ImplicitFreeSurface and ExplicitFreeSurface.) I'll open an issue on Oceananigans.jl soon.