Compressible Navier-Stokes on TreeMesh3D

[andrewwinters5000]

This implements the three dimensional compressible Navier-Stokes equations on Cartesian meshes for TreeMesh{3}. It was quite straightforward to extend the 2D methods for gradients and get things up an running in 3D. Below is a laundry list of what needs done before merging.

• Write tests for new elixirs, probably requires a new test/test_parabolic_3d.jl set
• Create a MMS convergence test that exercises the boundary conditions
• Get things up and running on DGMulti{3} as well
• [ ] Viscous shock tube test? Save for 1D Navier-Stokes, it would be a good student project
• [ ] Compute kinetic energy dissipation rate (as an analysis quantity) Compute in postprocessing from the integrated kinetic energy values and times available when analysis quantities are saved in an output file.
• Compute enstrophy (as an analysis quantity)

The points above address a few of the list items from #1147 .

[codecov]

Codecov Report

Merging #1239 (d7c31a7) into main (15e2627) will increase coverage by 0.16%.
The diff coverage is 94.90%.

❗ Current head d7c31a7 differs from pull request most recent head bc4382d. Consider uploading reports for the commit bc4382d to get more accurate results

@@            Coverage Diff             @@
##             main    #1239      +/-   ##
==========================================
+ Coverage   88.71%   88.87%   +0.16%     
==========================================
  Files         328      334       +6     
  Lines       25138    25785     +647     
==========================================
+ Hits        22300    22914     +614     
- Misses       2838     2871      +33     

Flag	Coverage Δ
unittests	88.87% <94.90%> (+0.16%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
src/Trixi.jl	70.00% <ø> (ø)
src/callbacks_step/analysis_dgmulti.jl	80.65% <0.00%> (-18.04%)	⬇️
src/equations/compressible_navier_stokes_2d.jl	98.96% <ø> (ø)
src/equations/equations_parabolic.jl	100.00% <ø> (ø)
src/solvers/dgsem_tree/dg.jl	100.00% <ø> (ø)
src/equations/compressible_euler_3d.jl	97.34% <83.33%> (-0.28%)	⬇️
src/solvers/dgsem_tree/dg_3d_parabolic.jl	94.72% <94.72%> (ø)
src/equations/compressible_navier_stokes_3d.jl	99.12% <99.12%> (ø)
...ples/dgmulti_3d/elixir_navierstokes_convergence.jl	100.00% <100.00%> (ø)
...ulti_3d/elixir_navierstokes_taylor_green_vortex.jl	100.00% <100.00%> (ø)
... and 5 more

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

[andrewwinters5000]

Two questions to @jlchan @ranocha @sloede:

1. The co-rotating Euler test failed due to some exception. Have y'all had this happen before?
2. The new parabolic tests all pass in 57m; should I modify them to be shorter to not extend the automatic testing?

[ranocha]

1. The co-rotating Euler test failed due to some exception. Have y'all had this happen before?

I haven't seen that before. I opened #1242 to test main.

2. The new parabolic tests all pass in 57m; should I modify them to be shorter to not extend the automatic testing?

No, we have some CI jobs running longer than that, so it's fine.

[andrewwinters5000]

I haven't seen that before. I opened #1242 to test main.

The test fails there as well with the same error.

[ranocha]

@sloede Any ideas?

[andrewwinters5000]

I have been trying to work out the best way to create an enstrophy callback for the 3D compressible Navier-Stokes. I have written the necessary routines that live in src/equations/compressible_navier_stokes_3d.jl with my thought process being that it can reuse the gradients already computed and stored in cache_parabolic. This, though, proves a bit unwieldy because the analyse and integrate routines are designed to only dispatch on the advective cache (within semidiscretization) so the callbacks do not have access to the cache_parabolic that I want access to. Do y'all have any good suggestions on how to give the integrate (or integrate_by_indices) routines access to the gradients information?

[jlchan]

This sounds similar to the issue we had with initial conditions and parameters that depend on the equation type.

I had experimented with a "combined" equation type for dispatch in #1160 but I think this would basically necessitate rewriting a bunch of the callback codes.

[jlchan]

What if we introduced a separate set of callbacks that dispatch on the viscous cache and viscous equations?

Alternatively...is reusing computed gradients worth it?

[ranocha]

You could try to intersect the call chain at

Trixi.jl/src/callbacks_step/analysis.jl

Line 462 in 4afc4b7

function analyze(quantity, du, u, t, semi::AbstractSemidiscretization)

There, the semidiscretization is passed as an argument - which should hopefully be the hyperbolic-parabolic semidiscretization in this case.

[andrewwinters5000]

Alternatively...is reusing computed gradients worth it?

I was hoping to reuse the gradients because then the callback would not be solver specific. If we re-compute the gradients on the fly, which I am open to doing, then the callback would (maybe) need two version; one for DGSEM and another for DGMulti

[jlchan]

Would we need two versions anyways since the interfaces are different? DGMulti accesses the array indices directly while TreeMesh, ... use accessor functions like get_node_vars?

If you see a way to define a unified callback, then we could probably use multiple dispatch to specialize on gradient calculation.

[andrewwinters5000]

@andrewwinters5000 mind if I mark this ready for review?

Sure that works for me. I am not sure why the coverage drops.

[jlchan]

I think it's a spurious drop. The coverage report is showing drops in areas that we didn't touch.

[sloede]

This is just the first part of the review, where I only looked at some small files. It would be great if no non-essential merges or edits were to be made to this PR (except for addressing the comments) until the review is finished, such that I do not lose track of which files I have already looked at 😬

[sloede]

I have made a final review and left a few comments. Besides that, I think we are good to merge this.

We should keep an eye on the duration of the parabolic tests in GitHub though. It seems like they are already taking quite a long time, thus if we're adding more systems/tests, we need to redistribute.

[ranocha]

I think this should be ready to merge

[jlchan]

LGTM - thanks @andrewwinters5000 for spearheading this!