Видео

Currently yes with paraview. I am working on a full scale simulation of a reactor like Helions. This will be ~10 billion cells though and require my own rendering engine due to the size. I need to update it but here is the current prototype for the rendering github.com/loliverhennigh/PhantomGaze

@@oliverhennigh451 Nice! Thinking of writing my own visualization for my kinetic solver too but for different reasons: no visualization software supports arbitrary order FE solutions on different bases (except GLVis). Also, its a lot of fun to do visualization and computer graphics!

I am still writing the solver. Not a whole lot yet but once I finish I will put it on GitHub.

silence, bot

@@kartoffelbrei8090 nuh uh

@@sciepan666 This is like exactly what a untrained LLM would produce. Wild stuff man

Just Euler equations and FV with explicit time stepping. It’s on a dense grid as well but don’t remember the resolution. I can get around 500 million cells on my 4090 and if I do some tricks can get up to 4 billion using out of core memory.

This was a fairly small simulation run on my RTX 4090. Sim size about 4 million. I will rewrite the solver in Taichi or Warp though and should be able to get ~250 million cell sims running.

@@oliverhennigh451 250 million is really nice! I would assume using something like Taichi etc. saves a bunch of time compared to CUDA/OpenCL in the development, but what about simulation speed? Do you know your achieved vs. theoretical throughput?

@@yunuszenichowski This is a good question that I would also like to know the answer to haha. I have done some pretty thorough benchmarks for Warp and CUDA/OpenCL on Lattice Boltzmann methods and found them achieving same performance. I would assume will also be the case for Taichi but should test it when I get a chance. LBM is a bit of a contrived problem though and its not terribly difficult to get optimal performance. I was hopping to get my solver in a better place and compare to Athena++ when I run on the CPU. Not sure how to compare the GPU though...

@@oliverhennigh451 The bare LBM implementation is relatively easy to make fast, but it obviously gets more difficult, the more features and methods you add. I really like the idea of abstracting away the performance optimization, since it's very tedious if you don't know what you're doing, like me :) . And you could get cross-platform, which I really miss using CUDA.

This is written in Python using JAX (github.com/google/jax). I wrote this in JAX to do automatic differentiation through the solver and allow for design optimization. You can see this here where I optimize the initial conditions so that at a given time the plasma forms an image of a smiling pumpkin ruclips.net/user/shorts7l-JdndaqBo. Ultimately I will rewrite this in Taichi or Warp which will allow for ~10x larger simulations. github.com/NVIDIA/warp www.taichi-lang.org/. I have an inviscous Euler solver that can run a 500 million cell sim on my 4090 and it uses Taichi.

@@oliverhennigh451 I probably discovered your channel from a post on the Taichi discord? I'm really interested in playing with Taichi. I was also interested in Warp (thought I saw it used for differential renderering/physics but have a very hard time finding documentation on it)

It took about two hours in paraview although I might try writing my own ray tracer now. The actual computation was around 3 hours though.

@@oliverhennigh451 wow you write your own? Cool! Which language you're using for ray tracer?

@@KangJangkrik I am using a library called warp, github.com/NVIDIA/warp. It lets you write optimized kernels in python. I only learned about it around 4 months ago and it has been awesome to use. Makes it really easy to implement optimized code that runs fast on GPUs. This is how I wrote the PIC solver and it only took a day or so to write (~400 lines of python). They have some ray tracing examples I was looking at and might give it a go.

@@oliverhennigh451 awesome! Mind to create tutorial video? :)

I’ll think about it :)

Why not?

Mostly for fun but also to show what is possible now. This simulation is actually about as big as the one here for example ruclips.net/video/lgxVYl_pslI/видео.html (their grid size is bigger). It only took me a few hours on a single gpu and can probably be made a bit faster/bigger.

Hey Mehras, I'm a bit surprised anyone sees my videos haha. Right now I am wrapping up this and some other stuff in a paper for ICRL so hopefully I will be done Oct 27th (fingers crossed). After that I am going to make a fun video explaining all the stuff in the paper. This work was actually supported by the AIgrant thing and I made a video outlining my stuff here (ruclips.net/video/3ZbmrUH9eto/видео.html). I'm about half way done with my project so my video will be like a what did I do so far thing. Its weird to watch that old video and realize that a lot of the stuff I ended up doing didn't work the way I thought it would haha. That damn unsupervised loss function idea I had took 2 months of my life and didn't work at all. Basically what is happening in the video is a network is predicting the steady state heat dissipation from a list of parameters corresponding to the height of the fin things. Then I kinda treat the heat at a loss function that I minimize by doing gradient decent on the parameter space. The fins take heights that try to cool off the heat sink as well as possible. Probably a bad explanation. The name of the video is a little miss leading too but I haven't really settled on a good way to say it. Here is a rough draft of the paper abstract if you are interested. I rewrite the whole thing every day so I'm not really sure if this is how it will come out but I will post a link to the paper when its finally done. " In this paper, we propose a novel method that makes use of deep neural networks and gradient decent to perform automated design on complex real world problems. Our approach works by training a neural network to mimic the fitness function of the optimization task and then, using the differential nature of the neural network, we perform gradient decent to maximize the fitness. We demonstrate this methods effectiveness by designing an optimized heat sink and both 2D and 3D wing foils that maximize the lift drag ratio under steady state flow conditions. We highlight that our method has two distinct benefits. First, evaluating the neural networks prediction of fitness can be orders of magnitude faster then simulating the system of interest. Second, using gradient decent allows the design space to be searched much more efficiently. " Oh ya, all the code is here (github.com/loliverhennigh/Flow-Sculpter) but it is super messy right now and I dont have time to clean it up till after the paper deadline. I do have some similar type projects like github.com/loliverhennigh/Computational-Fluid-Dynamics-Machine-Learning-Examples and github.com/loliverhennigh/Phy-Net.

I made a somewhat more detailed video ruclips.net/video/Xr91QAPBBa4/видео.html