Comment by thomasjames | Hacker News Reader

Comment by thomasjames | Hacker News Reader

Parent Full thread

The truth about all flow and especially turbulence is that it is not hard to make it look convincing to the human eye. Capturing the motion on various length scales and presenting it in a plausible way is not all that computationally intensive (the algorithms to make them can still of course be challenging). Making simulations that are numerically correct, however, is incredibly computational intensive (in addition to being algorithmically complex). This is why CFD is still run on massive clusters for transient simulations. There is also a whole field in itself for "sub-grid" turbulence models that allow for the spatial resolution of a simulation to remain coarser than the Kolmogorov microscales (https://en.wikipedia.org/wiki/Kolmogorov_microscales). (As a side note, Kolmogorov was a true polymath/genius in areas outside information theory and complexity theory.) Cutting edge basic research on turbulence will perform calculations at such scales (direct numerical simulation). Suffice it to say, most graphics work does not waste too many cycles on physical accuracy. Hard to know for sure, but since the video has words "cinema" and "rendering" in the description as opposed to "simulation", it is probably just realistic enough to pass muster in terms of human perception. Also accurately simulating the formation of foam requires a two-phase simulation of both gas and liquid as well as considering cohesive forces.

Small nitpick... my understanding is that these simulations are numerically correct- they are, however, numerical approximations to simplified versions of the Navier-Stokes equations. So the mathematical equations are incorrect in that they do not accurately model parts of physical phenomena they represent. They are accurate enough if your objective is to create pretty (moving) pictures. They are not accurate enough if your objective is to determine forces on objects.

Point well taken. They are not numerically unstable in the sense that the term is most often used with ill conditioned matrices or floating point weirdness, but yes, they are significantly oversimplified and will not converge with reality if you measure them by a meaningful metric.

Excuse my curiosity- do you work in CFD? I did quite a bit of CFD work for my degree- then I made the mistake of becoming a programmer :).

I worked for a CFD HPC company for about two years doing a mix of traditional engineering and programming, and I realized that I like programming and learning about computer architecture way more than the mechanics. So I went back to school for CS/CE and just like you fell into the trap!

Great, thanks for the reply.

I do hydrodynamic modelling at work but on much larger scales (hundreds of metres to tens of kilometres - with FVCOM (http://fvcom.smast.umassd.edu) in particular) and the turbulence is handled by sub grid scale parameterisations, like you say (in my case, the model used is GOTM: http://www.gotm.net). I wasn't sure if these renders were using the same techniques or were pretty animations which looked OK.

Small world, I used to work for an industrial CFD company, and one of my tasks was estimating the computational effort required for a given simulation (which depended on the specific characteristics of our turbulence model). I was mostly doing compressible work, too, so my knowledge of hydrodynamic models used in cinema is not super detailed. I do know, however, that most cinema software uses particle based simulation for incompressible flow, which really is not a terrible approximation, but it probably would not be suitable for the kind of work you are doing!

Reply on news.ycombinator.com