fluid simulation using SPH and CUDA

This is a movie from my thesis project. We are investigating interactive simulation of fluids on the GPU.

Some more details:
* timestep used: 0.008ms
* FPS: 30
* particles: 131k
* world size: 10x2x10
* uniform grid with bucket sorting

See http://vloeistofsimulatie.blogspot.com (dutch site)

Category:

Science & Technology

Tags:

• SPH
• CUDA
• fluid
• simulation

[NoctumDeVir]

HI!

That's nice!

Do you use particle-particle collision or just a standart SPH force method?

And have you verified the scale, needed to resolve the boundary layer?

If you take DNS scale as l=(nu^2/eps)^1/4 you get A LOT of particles =)

Thanks!

[KjelleApers]

Hi thanks,

We use SPH, not a mass spring system or something like that. So collisions are not calculated. I think it'll be expensive to take collisions in account... detecting them is easy but you need to react to them, for which you need to use a smaller timestep... that is not really parallellizable :-)

[KjelleApers]

For the scale... u can use some different methods to avoid needing a lower scale at the boundaries. A boundary can be represented as just an extra external force. This is used in the paper of Harada on sph on gpu's. Another way is by just doing simple physics, using the normal of the surface. You won't get nice turbulence results though ;-) SPH is just an approximation of fluid motion and suitable for realtime simulations.

I hope that answers ur questions, and good luck with your simulations!

[wi2d89]

hello, i'm Dian, i really interest with your thesis. Now i have the same topic for my Final year project. Can i have the source code ?thank you

[KjelleApers]

I am sorry, we can't give source code. Take a look at the particles code in the CUDA SDK and read the SPH fluids paper of Müller to get started. Our thesis was about combining height fields ands SPH, but is all in Dutch.

Have fun with you project

[NoctumDeVir]

@KjelleApers Thank you for the reply.

I'm implementing one tight away now =). But with some problems =(

I have some more questions. Would it be patient for you to answer, please =)

1. Do you use pressure gradient kernel like this (say on axis x): gradW_x(r,h)=-45/Pi/h^6*r_x/||­r||*(h-r)^2? And pressure force is Fp_x=0.5*sum[ (p_i+p_j)*mass/rho_j*gradW_x ]; i!=j.

2. Do you normalize all kernels like: sum(W(r,h))=1?

3. How many particles are usually interact within the radius of dependence (h)?