EB-inflow capability - ... problems #540
Description
I'm noticing some weird behavior on this experimental inflow on EB capability I thought others may be interested with. It's quite difficult to describe, and I have tried a few things to solve the "issues" already, but so far I'm at a loss. I thought I would try to describe what I'm seeing with images, and if anyone has any thoughts, I'm happy to hear them!
The case I'm trying to implement is quite simple in theory: 4 circular inlets on the side of a cylinder, with a few cells to discretize each one. I'm only pushing the inflow on a plane, say the XY one. I can align my inlets (that is, making paired of inlets face each other) or slightly off-center them and the results are similar.
In 2D, I see the expected flow patterns both on CPU and GPU, but as soon as I'm 3D I see some anomalies in the velocity right after the initialization, and these are much worse on the GPU and in the X-direction. I actually think there may be 2 problems.
On the CPU, it's the Z velo that gets initialized with a few non negligibly high values in a few cut cells, and it concerns the inflow aligned with the X-dir the most. Actually, I've changed the geometry to inflow on a square instead of a cylinder and it's much more obvious (see picture1 and picture2)
It creates an initial asymmetry which quickly dissolves as the computation resumes, and we eventually see the expected flow patterns for all inlets (see picture3).
On the GPU, I see the same issue with the Z velo, but my X and Y velocity on the inflow that is aligned with the X-dir gets initialized with extremely high values in a few cut cells. The Y dir inflow seems to be behaving much better, almost as expected (see picture4).
Now, the square duct geom on the GPU gives different problems, but still on the X-dir aligned inflow ... but this time, the X-vel is too low, and we end up with an asymmetrical pattern where my Y-dir jet progresses faster...
In other words: it's a mess.
List of things that I've tried:
- 2D
- making the body a square duct and not a cylinder (the inlets are still inflowing on a cylinder)
- centering the inlets (no difference)
- Aligning the geometry with the cells (volfrac either 0 or 1) --> and this solved the issue both on CPU and GPU, everything is symmetrical, min/max of X and Y velo are the same, jets penetrate the same way etc. (see last picture)
- but then playing with the redistribution scheme made no difference ...
I'm happy to provide more info to help in the discussion; at this time I'm not exactly sure what else to try or where the problem could come from (is it a Pele or an AMReX thing?). At the very least: you have been informed!