> Thanks for the explanation, i was more referring to the actual program
> though.  The nice thing about particle based fluid simulation is that
> the fluid can move in large areas without excessive memory requirements
> that would occur with grid based methods.  It could even be worth
> considering integrating such a system into POV since for interaction
> with the environment you could rely on the raytracing functions.

In short, there are two different computational phases:

1) Compute change rates for every particle involved in the simulation. In detail:
    -) fluid-fluid particles interactions (every particle within a fixed distance from
        the particle being updated has to be considered in the process. Check
        the papers for more details about the equations... you have to update
        accel. and density. ~30-40 neighbours / particle).
    -) external forces (gravity... simply add them to the accel. vector of the
particle).
    -) boundary-fluid particles interactions (for the moment boundary is
        modelled introducing fixed particles in correspondence of the solid
        surface and adopting a pure repulsive Lennar-Jones interaction force.
        Fixed particles are placed really close in order to form a strong
        barrier to the penetration of the fluid). It should be easy to replace
        particles and to manage directly particles-solid collisions.

Note: 1a) In order to speed up computations, as the interpolating kernel
          presents a compact support (only particles within a fixed distance
          interact), it could be used (well, it's obligatory ;) ) a suitable
neighbours
          finding algorithm. I've used a simple uniform grid: all the
          neighbours of the particle being updated are located checking only
          few adiacent cells (you can take advantage of the simmetry in
          particle-particle interactions to reduce computations).
          1b) Current time-step is determined checking the physical properties
          associated to every particle (simply check accel. and speed of sound).

2) Update particles position-density-speed (leapfrog scheme, prediction - correction
     formulation) using change rates and timestep found in the previous step.

The first one is the more complex and time consuming phase by far...
1661s computation -> 1558s phase 1 vs 102s phase 2... ;)

> You can use the density file in an isosurface shape just like any other
> pattern:
>
> #declare Fn_Density=
>    function { pattern { density_file df3 "file" interpolate 2 } }
>
> isosurface {
>    function { Fn_Density(x, y, z) - 0.5 }
>    ...
> }
>
> This will show the isosurface with density=0.5.

Gotta try it but I have to rerun simulation to collect data once again, though.
I'll do it in the weekend.

Post a reply to this message