Stiff Deformable Materials in Realtime

[madMAx43v3r]

Hi,

i want to simulate stiff materials in realtime with a spring-mass network.
first i had no clue about numerical integration.
so i started with simple forward euler, then found Runge-Kutta RK4 and eventually adaptive Runge-Kutta order 4 and 5 (RK45)
with Cash-Carp coefficients.
even with RK45 i didn't hit realtime, so i moved the whole thing to the GPU, only to find out that i was latency bound due to the high number of iterations. only with node numbers of 100k and more i was able to use full memory bandwith (around 100GB/s on a GTX260), but not in realtime of course.

so it seems i need to use implicit integration after all
i already read a lot of papers, but people always talk about constraints, not springs.

so my question is, can i use real implicit integration on a spring-mass network?

[h4tt3n]

Why exactly can't you do this mass-spring mesh in realtime using explicit integration? Any relatively new machine should be able to handle 10.000's of springs without sweating.

It seems to me like you're overdoing the integration part a bit. Did you try the relatively simple velocity verlet algorithm?

[madMAx43v3r]

im using spring stiffnes values above 100k, which does not run in realtime for about 1000 nodes with adaptive RK45CK...

but i just read that paper: http://www.uni-weimar.de/~caw/papers/p7 ... alesic.pdf
and it uses a verlet method like you said, so i'll try this
but it does not seem to be velocity verlet, because its using the midpoint velocity for spring damping calculations...
i thought damping (on velocity) is not possible with velocity verlet?

thanks so far

[h4tt3n]

Below is a link to a thread, where I recently described my approach, which I call an "iterative" velocity verlet algorithm. It'll handle spring stiffnes values of several millions (for particle mass 1 and timestep 0.01) with a reasonable number of iterations. It will definitely do what you requiest in real-time. Check it out

Btw. the method is very similar to what's described in the paper you link to.

http://bulletphysics.org/Bullet/phpBB3/ ... 3&start=30

Cheers,
Mike