Hacker News new | past | comments | ask | show | jobs | submit login

Sure. It's a correct implementation of the projection method [1] for solving the 2D incompressible Navier-Stokes equations, so it is in fact using genuine physics. In terms of physical accuracy, there are some caveats - it's using the Jacobi method (chosen for its simplicity) to iteratively solve the diffusion and pressure Poisson equations, so the accuracy will be determined by how many iterations you're willing to do. That approach won't scale to larger systems. There are also some limitations from the choice of collocated (as opposed to staggered) grids and the approximations used for the boundary conditions. Nevertheless, for a small 2D simulation, you can get physically accurate results as long as you use enough Jacobi iterations and have a sufficiently fine grid.

[1] https://en.wikipedia.org/wiki/Projection_method_(fluid_dynam...




Also it uses backward differencing, it makes it stable at long timesteps. That is great to run fast for visualization, but it is horribly inaccurate (or possibly even plain wrong) if your system has large gradients in flow speed.


Yes, I've found that computer graphics focused fluid simulations frequently choose stability over physical accuracy. These choices also result in unphysically high numerical viscosity. (I didn't check what finite difference stencil or finite volume scheme the code uses, though I presume it's a lower order accurate one that probably has a fair amount of numerical viscosity.) In principle if you reduce the grid spacing and time step it will converge provided the software doesn't use any tricks like approximate square roots, etc.

The numerical methods for solving the PDEs that computer graphics folks use would surely be considered primitive by someone who develops engineering computational fluid dynamics softwares.


> I didn't check what finite difference stencil or finite volume scheme the code uses, though I presume it's a lower order accurate one that probably has a fair amount of numerical viscosity.

It's just the basic second order central difference. It also uses a first order approximation to the Dirichlet and Neumann boundaries, so that additional error will diffuse throughout the simulation region. It doesn't use any approximation tricks for square roots etc., so given appropriate floating point semantics (as a physicist I have no clue what the shader language specifies there) you can still get realistic and accurate results by reducing the spatial and time steps, which is easily doable for a small 2D simulation on modern GPUs.

Ultimately, though, all the basic "best practices" for simulations of this kind - staggered grids, higher order derivative approximations, etc. - aren't very complicated and are well described in any CFD textbook. What really makes engineering CFD software complicated are things like handling complex geometries with dynamically refined meshes, efficiently solving the resulting linear equation systems at scale and coupling the fluid dynamics to other physical phenomena while retaining numerical stability and accuracy.


Don't get me wrong: Doing all the approximations to get a fast, visually appealing animation is perfect for what this demo aims to do.

Fluid simulations for science or engineering (think airflow around the next Boeing design or simulations of planet formation) are still very hard. And that is not because physicists are stupid, bad at coding or easily replaced with machine learning.


Sorry, I was unclear. I agree that the demo does what it intended to do. I was just listing a few additional reasons to believe it may not be physically accurate as-is. I know from previous discussions on HN that many readers are interested in this.


Totally good idea to do that. I just want other readers to be aware for what applications that is useful and for what applications other techniques are more appropriate.




Consider applying for YC's W25 batch! Applications are open till Nov 12.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: