Netgen

[davidcortesortuno]

Hi, I noticed that the pixi branch was merged into the main one. In this version the mesh utils were updated by Codex, but it seems most of the code is unnecessary, as the netgen binary is not being called, which requires some functions to address issues with netgen's GUI. Rather, the pixi branch uses Python's netgen 6.2, which seems not to have any of these problems.

I will see if I can simplify most of the mesh functionality. Something else we noticed with my group is that the maxh argument works differently in old finmag's netgen 5.3., which produces finer meshes compared to the new finmag, using the same max value.

[fangohr]

Hi David, best to use the code with tag python2 for now for the production work in your group. This is the old master, and not changed. Sam and I have more changes to be merged into main soon (i.e. this week) - best not to make any changes before that is merged.

It is possible that some of the CI jobs use Netgen in 'unusual ways' to provide comparisons for the porting of the code. We can review and remove that when all our porting is merged to main.

Your input on netgen's behaviour is welcome, and we should update the code on main eventually.

[davidcortesortuno]

Ah, good to know the old finmag is tagged. I'll wait for the updates then and check netgen's functionality in the meantime.

[samjrholt]

Hi David! The attempt to restore the basic master functionality with dolfinX is on the dolfinx-parity branch (merging to main soon):

On the unnecessary netgen-binary code: this resolves itself with the merge. The mesh layer was rebuilt during the port - mesh generation now goes through Gmsh/OCC (maxh maps to Gmsh's Mesh.MeshSizeMax), the legacy Netgen-CSG .geo text is parsed directly and retained only for caching/traceability, and there are no netgen binary calls or GUI workarounds left anywhere (netgen_is_usable() returns False by design; the binary backend was dropped for now in an attempt to get the basics working).

On maxh producing coarser meshes than netgen 5.3: I saw this as well and it persists under the new generator too. We measured exactly this while investigating a comparison-tolerance question: regenerating a mesh from the same .geo at the same maxh gives 1202 nodes / 4960 tets today vs 1537 / 6886 in the archived netgen-5.3-era run — ~25% fewer nodes.

[davidcortesortuno]

Hi Sam. Great, I'm looking forward to these changes.

I was delving into this issue with Claude as well, and Netgen has another mesh partitioning tuning parameter, curvaturesafety which can be increased from 4 to 6 or so, which helps to improve the mesh resolution in curved geometries.

I see gmsh is now the preferred mesher, but I guess we could still leave Python netgen? Since the pixi branch included this package in the environment. I remember netgen has decent optimisations for tetrahedralizations. And the geometry geo definitions are still there.

[fangohr]

@davidcortesortuno - I updated the README.md to explain the story. Some of the tags do not exist yet, but Sam will create them when the outstanding pull requests are merged.

[samjrholt]

@davidcortesortuno I have merged a working state into main. This is not performance optimised but most of the functionality is transferred. Details are available in dolfinx-transition.md. What is already supported is documented in docs/SUPPORTED.md. I will take a pause on some of the development now so please feel free to go ahead and test it and make edits!

We have purposefully got AI to leave a significant number of comments in all of the files in order to give us (and other AIs) better context when trying to edit. I have left these in for now whilst we continue development but at some stage we should do a clean up.

[davidcortesortuno]

@samjrholt Awesome. It's a long list of changes. I'll have a look of what can be done.
Regarding the mesh, we can add optional optimisations like gmsh.model.mesh.optimize("Netgen").