VDB boundaries colliding on FLIP

[caskal]

Hey guys,

I'm having some issues with a FLIP collision, seems my VDB collider boundaries are affecting the sim, anyone knows what can be causing this? 

I'm using static object + static solver with a proxy volume as collider.

Thanks!

 

[Omni]

Faced the same thing before with FLIP, once again right now using grains. I'm trying to figure out what the solvers are doing(either FLIP or POPs for that matter) in that case. It seems that feeding a moving collider that was pre-baked via Volume Sample is the issue. If i use Ray Intersect it's fine but it ruins the purpose of optimizing your collisions for sim times. As far as i can tell, it's as if the solvers are(not?) performing a volumesample function on the static object for the collider and the points outside the bounds of the volume are getting a value of zero?..Only of the edges of the volume bounds...? Which would result in a collision surface being detected...

Again, using ray intersects works so it's really confusing to tell why the particles collision interpretation changes between those input methods...

[Omni]

The SideFX documentation for the Static Object DOP says this about Volume Sample method:

 

Volume Sample

The divisions are ignored in this mode, instead they are computed from the first volume or VDB primitive in the geometry. The computed divisions are chosen to match the voxel size of the source volume. The volume primitive is sampled raw and treated as a signed distance field. The assumption is that the source is the output of an Iso Offset or VDB From Polygons SOP. If it isn’t a true signed distance fields, unusual things may happen with RBD collisions.

 

In my case i'm always using the vdb from polygons method so i guess it should work as intended. I'm aware this reply doesn't help much but it felt it makes sense to say that Houdini expects the Volume Sample to be fed exactly this...

[Omni]

After trying a few approaches this morning, a fix that worked for me in the case of POP grains was to use the exact same voxel size for the collider as the particle separation size of the POP grains microsolver.

 

Hope that helps anyone, that might as well be a clue for the FLIP solver...

[sebkaine]

i get the exact same issue, but i can't remember having seen this before , weird ?

- i use pre cache vdb collider loaded as static object 

- the collider is a deformed object

- when i don't activate viscosity everything work smoothly 

- as soon as i activate viscosity the solver stop to use the vdb collider, and is unable to evaluate the sdf grid properly

- instead it use the bouding box of the collider 

I'll investigate on this and this if i can find a proper solve !

Cheers

 

I have attach a simple scene that illustrate the problem.

 

flip_collision_problem.hiplc

Edited November 14, 2021 by sebkaine

[ryew]

Which version are you running in?  I don't see the issue in either setup using H18.5.499, just interaction immediately around the torus surface

[sebkaine]

thanks for taking the time to open the scene ryew. i'm on 18.5.633.

the weirdest thing is that after your comment,  i just reopen the hip delete the noViscosity group and test the viscosity only, and now it looks to works. 

ahaha something esotherical here, generally in those situation the problem come from behind the keyboard.

the hip i share is a quick exemple of a scene i can't share here, where the problem is the same. i'll double check that tomorrow. 

Cheers

E

[sebkaine]

well i think i get where the problem come from. 

my collider was not properly setup. I use direct vdb conversion instead of the collisionsource SOP.

thus my emitter wasn't properly evaluating animation on the collider for substeping.

things works without viscosity and stick on collision because i guess houdini only need one substep to solve the calculation.

but when visco and sticky collision were activated houdini need more substep.

so i think that each time the solver was trying to substep , it can't find proper info from the collider and thus fallback on the bbox of the object. 

using collisionsource that use a proper timeblend solve the issue. 

 

Edited November 15, 2021 by sebkaine