LAPD solver in 3D

[jwscook]

DM has just fixed the outflow of the toy 1D model. This fix is being applied up the chain of codes and models.

[jwscook]

Owen has found a problem in DM's 1D outflow - need to ask him this friday.
Ed to recheck that Firedrake implementation works with Mach number of 0.1

[jwscook]

Ed has something going in FireDrake with sonic outflow

[jwscook]

We still need Dave's help to get the NESO version working for the subsonic outflow case.

[jwscook]

Dave is now back need to make a prioritised list:

1 DG or CG explicit time anisotropic diffusion (currently all examples use implicit timesteps e.g. https://github.com/ExCALIBUR-NEPTUNE/nektar-diffusion/blob/rfu_ambipolar/example/Unsteady_ring/ambipol.xml)
2 this sub-sonic outflow
3 nektar-waveeqn for implicit steps.

[jwscook]

Back now as priority

[jwscook]

Work has started in FireDrake over here

We need to decide on CG vs DG implementations where the dynamics aren't damped too quickly i.e. keep going as long as the Hermes3 simulations go.

[jwscook]

Using Rogers & Ricci 2010 in Firedrake

[jwscook]

Details on normalisation https://github.com/ExCALIBUR-NEPTUNE/firedrake-lapd/blob/rogers-ricci_v1/docs/rogers-ricci.md

[jwscook]

Current activity is checking units and both @oparry-ukaea 's and @ethrelfall 's firedrake implementations for consistency.

[jwscook]

3D runs with 1:1 mass ratio - investigating this to move to 1836 (R&R only use 400 in their paper).

[jwscook]

3D is still a bit crashy. Jack's speed ups gives 20% improvement on the 2D version. 2D doesn't crash.

• The 3D solution diverges when running with more than 1 node on Archer2 < big problem
• 32 CPUs takes 15 hrs to get the crash point < way too slow.
• Try periodic boundary conditions to see if it's the parallel +/- sound speed BCs causing the slow down.
• Try turning on solver iteration counts via PETSc's logger

[jwscook]

Struggling with proper BCs for electron velocity, https://github.com/ExCALIBUR-NEPTUNE/firedrake-lapd/blob/rogers-ricci_v1/docs/rogers-ricci.md#boundary-conditions: not sure if BCs are applied in Firedrake.

These BCs are a function of electron temperature in time, so one needs to apply this as an EquationBC https://www.firedrakeproject.org/firedrake.html#firedrake.bcs.EquationBC?

[jwscook]

The new BC fixes the problem is ~10x slower than the version w/o BC.

[jwscook]

Made simplifications to the 3D model but it's still not working despite the potential solver being the same as in 2D.

• make sure the potential normalisation is adequate: $\phi \approx 1$.
• loop in Romain to inspect the fine grained output
• loop in the Firedrake developers on an MWE

[jwscook]

This has morphed into a Rogers & Ricci solver...

[jwscook]

Testing BCs in true 1D in Firedrake (Ed) and in Nektar (Owen) to then move to 3D.

[jwscook]

Using 1D-in-3D version of equations to make finding the indices on trace points on boundaries more general so that the solver works in 3D for any number of expansion modes / mesh / MPI ranks.

[jwscook]

Waiting for DM to implement the non-isothermal form in 1D. Difficulty writing non-conservative equations into Nektar++. Ed working on 3D version in Firedrake.

[jwscook]

Still waiting

[jwscook]

It works in 1D in Firedrake - not tried in 3D yet. Still waiting on DM to implement 1D version in conservative form.

Generally speaking, it is hard to solve non-conservative form equations in Nektar++'s DG solver.