MLEBABecLap aniso 2D and 3D

[eebasso]

Summary

Changes in MLEBABecLap for 3D and 2D to way flux is computed for EB surfaces.

The goal is removing the assumption that dx=dy=dz (anisotropic cells).

Additional background

The changes concern the way how Feb is computed. The changes in #2640 were only for 2D. This handles both 2D and 3D, with 2D changed to be more consistent with 3D.

Checklist

The proposed changes:

• fix a bug or incorrect behavior in AMReX
• add new capabilities to AMReX
• changes answers in the test suite to more than roundoff level
• are likely to significantly affect the results of downstream AMReX users
• include documentation in the code and/or rst files, if appropriate

[WeiqunZhang]

I only checked changes in 2D mlebabeclap_adotx_centroid. It does not seem correct to me. The grad_eb_of_phi_on_centroids_extdir function uses anrmx and anrmy as follows:

Real dphidn = rhs(1)*nrmx + rhs(2)*nrmy;
return dphidn;

If the return value is scaled by a constant, it can be re-scaled later. But in your change, the ratio of nrmx to nrmy has changed. So I don't see how that can be correct.

[eebasso]

In grad_eb_of_phi_on_centroids_extdir, rhs is the difference in phi in the x and y directions, but it has not been divided by a physical dx quantity yet. In other words, I am treating rhs and the output of grad_eb_of_phi_on_centroids_extdir as

rhs(1) = $\Delta_{x} \phi$ = dx $\frac{\partial\phi}{\partial x}$
rhs(2) = $\Delta_{y} \phi$ = dy $\frac{\partial\phi}{\partial y}$
grad_eb_of_phi_on_centroids_extdir = nrmx * dx * $\frac{\partial\phi}{\partial x}$ + nrmy * dy * $\frac{\partial\phi}{\partial y}$

This is why I scaled anrmx and anrmy the way I did. With anrmx = (apxm - apxp) dhx and anrmy = (apym - apyp) dhy, feb evaluates to

feb = beb * ( (apxm - apxp) dhx dx $\frac{\partial\phi}{\partial x}$ + (apym - apyp) dhy dy $\frac{\partial\phi}{\partial y}$ )

Apply definitions of dhx and dhy:

feb = B * beb * ( $\frac{\mathrm{apxm} - \mathrm{apxp}}{\mathrm{dx}}$ $\frac{\partial\phi}{\partial x}$ + $\frac{\mathrm{apym} - \mathrm{apyp}}{\mathrm{dy}}$ $\frac{\partial\phi}{\partial y}$)

[WeiqunZhang]

This handles both 2D and 3D, with 2D changed to be more consistent with 3D.

So the proposed change in 2D is not just more consistency, right?

[eebasso]

Yes, I think you are right, it's more than just consistency. The extra dxi factor is necessary to actually get spatial derivatives. This effect only matters when dx =/= dy. My apologizes for not adding a test to this PR. I can convert to draft if that's more appropriate.

[WeiqunZhang]

No problems. I thought the 2D was just a programming style change and I was focusing on whether the 2D behavior has changed.

[WeiqunZhang]

A number of regression tests failed with NaNs. Here is an example in https://github.com/AMReX-Fluids/incflo. It needs amrex and https://github.com/AMReX-Fluids/AMReX-Hydro to build.

$ cd incflo/test_3d
$ make -j TEST=TRUE
$ ./incflo3d.gnu.TEST.MPI.EB.ex benchmark.channel_cylinder-x  amrex.fpe_trap_invalid=1

It will crash. This test runs fine with the development branch. This does not automatically mean the issue is in this PR. But we need to figure it out.

[eebasso]

Okay, I will try updating the PR to the latest version of development to see if that helps.