2. remove all cells where a single point protrudes from the STL; smooth the surface to get a 'potato'; then, extrude cells back to STL. works flawlessly for convex geometry (a.k.a. sphere) but extrusion fails in concave corners and it's totally miserable when it has to snap to features (edges, vertices)

I tried these two with like 5 different iterations each. Was excited at first to see Claude do what I wanted but meshes were... shit

3. The verdict: snapping algorithm is kind of a gem (handles all possible situations with surfaces/features) and let's reuse that but try to make background more conforming/castellation better suited to the geometry. Alas, I can't describe what to do because I currently don't know how I'd do that.

Also: I think layer addition is not a problem - it just doesn't have space to work with. If the underlying mesh was pretty (like, a single thick cell layer) it would be a piece of cake - could be as easy as

refineWallLayer

...

And then I hit the weekly token limit. Now my floating license is used in the finance department where they're asking Claude about the weather

To avoid this there is @otaolafr's dream of creating layers based on a surface mesh first, and then fill the volume with cells (could be hex cells with polys to join volume mesh and boundary layer)

Replace block mesh instead of replacing snappy

Although maybe you'd run into the problem of snappy seeming to really hate cells that aren't perfect cubes

Engys advanced mesh algorithm sounds kinda like a fancy "buffer layer + refineWallLayer" strategy. Maybe you should ask Claude about it, just in case Anthropic scrapped that too 🙄 😅

1. For me, remeshing of surface mesh to quad mesh (if necessary) there are a bunch of existing algorithms that can deal with high complex geometries. 2. Extrude big feature conserving layer, Check the image of the code here https://hal.science/hal-03127350/document 3. Your main issue ones this is done is make conformal the big layer and the conformal interface now that's an issue, the fastest and at least doable thing to do is throw an advancing front tetra mesher. This can be netgen for example but there is a bunch of other meshers around. 4. You can divide you wall layer 5. You try to create a dual mesh that respects features, also difficult step. 6. I would (and if you want to ask to develop this as a extra tool, would be awsome) decompose the polys of the snappy refinement into "classic cells" tetra pyr prism blablabla. And use a smoothing based on the paper of 2. To increase the global quality.

This are the improvements I see. The hard point is 3. IMO or once you have you big layer your mesh snapped to to re create the big layer going the other way around to get conformal.

It has been almost fully solved by the snappy creator as he works in helyx. There you get way better quality with helyxmesher and it is based on snappy so you can have hopes that the truth is out there

for 2. also you have https://www.sciencedirect.com/science/article/pii/S0021999125000270 and if you add the 1. (before you say you want all full hexas) no, it is not the reason the reason is this: https://github.com/mxncr/AllHexBoundaryLayerTopologySolver

you have a bunch of material to throw at claude ^^

are you intentionally not using it for classy to keep it ai free? or simply you are not looking in that right now?

two last items, there is the all forgotten openfoam mesher with gui i can not find it right now that was in serial that first generated the boundary layer and it was generating quite good quality meshes... but in serial, also a place to check out.

and lastly there is a similar snappy boundary layer extrusion in salome cfd

oh and make claud throw mesquite before layer addition after snappying (or just before it) this will solve already a buuuuunch of issues if you success to remove the polyhedral cells.

i dont have the brain or the computer capabilities to do this. also if you have the time and want to help me there is 2 bugs in salome that it has been several versions that i am asking for help without solution, and if you can chase them with claud i would be infinity greatfull ahahah

(just if you simply want to make claud license work )

I dealt with custom backgrounds and the problem was that you can't refine a lot with that. after refining, points should be re-snapped to the surface or else you get dangling cells that make the situation worse than with non-fitted background https://cdn.discordapp.com/attachments/1547237441067294780/1547485271220551680/24846.png?ex=6aa39777&is=6aa245f7&hm=f85b5464fcf7b2d0d9b9f48c88dc69b4cacc11ef40142c41a49dfb0e97e48c09&

Besides, making a custom background is almost the same as making a proper hex mesh from the start, just a bit simpler. And it can't be smoothed or smashed into conformal-ish shapes, there's always something that goes wrong (spoils the mesh) or doesn't help.

Maybe an out-of-the-box approach: start with a tet mesh; extrude one wall layer, triangular; recombine it into quads and optimize; make a dual mesh inside; leave tets where it would create crap

tetra with prismatic layer?

why not use salome, it exists already it is there and it does it quite well.

the prismatic layer it is not an external package i think. but what you are saying with netgen is doable

i want to go back to work on this kind of things but i have my hole can of worms to deal with right now... 🥲 already i passed from my code taking +500gb and 1 week to 40 gb and ~2 days hahaahaha

that's the next plan! let's make a survey where cfMesh needs improvements

easy pease, add the buffer layer of snappy to it.

the buffer layer looks like is the last ultimate development in octree meshes to improve quality (and it is logic, it is dealing with missaligment which is the shitty thing of octree meshers)

or foamyHexMesh ahahahaha

lol

here you have another work on octree meshes https://internationalmeshingroundtable.com/assets/papers/2025/1017-compressed.pdf

that looked promising

When you see things like snappy and variations such as the Helyx mesh or cfMesh, it feels like it should be possible to fix the main issues of snappy. I really want to believe in the buffer layer thing. Tobias comments few months ago gave me hopes for some news in v2612

in his examples (true they are simple) but the minimum SJ is 0.2, openfoam is a piece of cake dealing with hexas of SJ 0.2 so you might be able to get better meshes but with this approach you at least will get meshes that will not make OF explode anymore for sure....

(and you get the extra benefit that you are full hexa)

if you fix the refinement to constant level in surface and features (not saying doing it but as a first approach) it would be a copy paste logic (so claud will handly pretty easily IMO)