Dr. Zsolnai covers a new turbulence simulation method that beats the 2008 Wavelet Turbulence standard. It upsamples across space and time, runs 5 to 8 times faster than high-resolution reference simulations, and compresses output to 600 times smaller with barely any visual loss.
Dear fellow scholars, this is Two Minute Papers with Dr. Zsolnai. Today is going to be all about turbulence. We are going to do this and we are going to do this and this and this. To understand what this new paper has to offer, we have to look at an earlier work named Wavelet Turbulence. Apart from the fact that in my opinion, it is one of the best papers ever written, it could do one thing extremely well. In goes a coarse, low resolution smoke or fluid simulation and out comes a proper simulation with a lot more details. And it is typically accurate enough to fool us. And all this was possible in 2008. Kind of boggles the mind. So much so that this work even won a Technical Oscar award. Please remember this accuracy statement as it will matter a great deal. Once again, it is accurate enough to fool us. Now, more than a decade later, here we are. This new paper is leaps and bounds better and can do five amazing things that the previous methods couldn't do. One, it can do spatiotemporal upsampling. Upsampling both in space and in time. What does this mean? It means that in goes a choppy, low resolution simulation and out comes a smooth, detailed simulation. Wow. Now that is incredible. It is really able to fill in the information not only in space, but in time too. So good. Now, two previous methods typically try to take this coarse input simulation and add something to it. But not this new one. No, no. This new method creates a fundamentally new simulation from it. Just look here. It didn't just add a few more details to the input simulation. This is a completely new work. That is quite a difference. Three, I mentioned that wavelet turbulence was accurate enough to fool us, so I wonder how accurate this new method is. Well, that's what we are here for. So let's see together. Here comes the choppy input simulation. And here is an upsampling technique from a bit more than a year ago. Well, better. Yes, but the output is not smooth. I would characterize it more as less choppy. And let's see, the new method, can it do any better? My goodness, look at that. That is a smooth and creamy animation with tons of details. Now let's pop the question. How does it compare to the reference simulation? Reality, if you will. What? I cannot believe my eyes. I can't tell the difference at all. So this new technique is not just accurate enough to fool the human eye. This is accurate enough to stand up to the real high resolution simulation. And all this improvement in just one year. What a time to be alive. But wait a second. Does this Even make sense. If we have the real reference simulation here, why do we need the upsampling technique? Why not just use the reference? Well, it makes sense. The plan is that we only need to compute the cheap coarse simulation upsample it quickly and hope that it is as good as the reference simulation, which takes a great deal longer. Well, ok, but how much longer? Now hold on to your papers and let's see the results. And yes, this is about five to eight times faster than creating the high resolution simulation we compared to. Which is absolutely amazing. Especially that it was created with a modern blazing fast reference simulator that runs on your graphics card. But wait, remember I promised you five advantages, not three. So what are the remaining two? Well, four it can perform compression, meaning that the output simulation will take up 600 times less data on our disk. That is insanity. So how much worse is the simulation stored this way? My goodness, look at that. It looks nearly the same as the original one. Wow. And we are still not done yet? Not even close. 5. It also works on super high Reynolds numbers. If we have some of the more difficult cases which where there is tons of turbulence, it still works really well. This typically gives a lot of trouble to previous techniques. Now one more important thing. Views are of course not everything. However I couldn't not notice that this work was only seen by 127 people. Yes, I am not kidding 127 people. This is why I am worried that if we don't talk about it here on two minute papers, almost no one will talk about it. And these works are so good, people have to know. Thank you very much for watching this and let's spread the word together. Perceptilabs is a visual API for TensorFlow carefully designed to make machine learning as intuitive as possible. This gives you a faster way to build out models with more transparency into how your model is architected and how it performs and how to debug it. And it even generates visualizations for all the model variables and gives you recommendations both during modeling and training and does all this automatically. I only wish I had a tool like this when I was working on my neural networks during my PhD years. Visit perceptilabs.com papers and start using their system for free today. Our thanks to Perceptilabs for their support and for helping us make better videos for you. Thanks for watching and for your generous support and I see you next time.
