Károly Zsolnai-Fehér explains how neural networks add fine splash and droplet detail to fluid simulations. A new ML-augmented pipeline called ML Flip reproduces high-fidelity movie-quality results while running roughly 10 times faster than traditional methods.
Fluid Implicit Particle.
Károly Zsolnai-Fehér.
March 17, 2019.
Approximately 10× speedup.
Dear fellow scholars, this is two Minute papers with zsolnai. If you have been watching this series for a while, you know that I am completely addicted to fluid simulations. So it is now time for a new fluid paper and by the end of this video I hope you will be addicted too. If we create a virtual world with a solid block and use our knowledge from physics to implement the laws of fluid dynamics, this solid block will indeed start behaving like a fluid. A baseline simulation technique for this will be referred to as FLIP in the videos that you see here and it stands for Fluid Implicit Particle. These simulations are often being used in the video game industry in movies, and of course I cannot resist to put some of them in my papers as test scenes as well. In games we are typically looking for real time simulations and in this case we can only get a relatively coarse resolution simulation that lacks fine details such as droplet formation and splashing. For movies we want the highest fidelity simulation possible with honeycoiling, two way interaction with other objects, wet sand simulations and all of those goodies. However, these all take forever to compute. This is the bane of fluid simulators. We have talked about a few earlier works that try to learn these laws via a neural network by feeding them a ton of video footage of these phenomena. This is absolutely amazing and is a true game changer for learning based techniques. So why is that? Well, up until a few years ago, whenever we had a problem that was near impossible to solve with traditional techniques, we often reached out to a neural network or some other learning algorithm to solve it, often with success. However, it is not the case here. Something has changed. What has changed is that we can already solve these problems, but we can still make use of a neural network because it can help us with something that we can already do, but it does it faster and easier. However, some of these techniques for fluids are not yet as accurate as we would like and therefore haven't yet seen widespread adoption. So here's an incredible idea. Why not compute a course simulation quickly that surely adheres to the laws of physics, and then fill the remaining details with a neural network? Again, FLIP is the baseline handcrafted technique and you can see how the neural network infused simulation program on the left by the name ML Flip introduces these amazing details and if we compare the results with the reference simulation, which took forever, you can see that it is quite similar and it indeed fills in the right kind of details. In case you are wondering about the training data it learned, the concept of splashes and droplets flying about, you guessed it right. By looking at splashes and droplets flying about. So now we know that it's quite accurate. And now the ultimate question is how fast is it? Well get this, we can expect a 10 times speedup from this. So this basically means that for every 10 all nighters I have to wait for my simulations I only have to wait one. And if something took only a few seconds, it now may be close to real time with this kind of visual fidelity. You know what, sign me up. This video has been kindly supported by my friends at ARM Research. Make sure to check them out through the link in the video description. Thanks for watching and for your generous support and I'll see you next time.
