Fluid Simulation for Graphics on Non-Uniform Structured Grids

I've been playing around with fluid simulation on non-uniform structured grids.  They have some charms in that it is easy to have very large domains with isolated regions of interest; e.g. near the camera or salient fluid features.  One of the big advantages is that it makes far-field boundaries easy, you simply extend the mesh far away from the domain.

My solver pretty run-of-the-mill; only the pressure solver was updated in order to handle anisotropic cells.  I've used a finite-volume formulation to derive the pressure-correction equation for this case, but it suffices to say that only the weights used in forming the Poisson equation change.

Here is an example, a grid that is roughly 10x10x5 meters, with 2cm fine cells for roughly 100x100x100 cells:

The aspect ratios get quite high, 100:1 is not uncommon.   You can see a buoyant flow simulation that I'm running as the contour values.  The simulation takes about 20 seconds per frame, of which 7 seconds is writing the VTK output files that I use to post-process.  In spite of that, the visual detail in the region of interest is excellent (given the resolution):

I find that it helps to advect a uniform resolution density field that is roughly 2X the fine cell resolution.  Writing the code with general-purpose fields that can be point-sampled arbitrarily makes this a trivial feature to implement, simply allocate different grids for density and velocity.

Finally here is a rendering of the final simulation in Paraview, total simulation time ~35 minutes. There are still some bugs to track down, but the results are pretty promising: