Google Brain and NYU trained a neural network on fluid and smoke simulations so it can predict how smoke moves in milliseconds instead of minutes. The network matches the quality of traditional simulators but runs far faster, though it's harder to validate than physics-based methods.
Dear fellow scholars, this is two minute papers with this piece of work is still in progress done by one of the members of the Google Brain research team and several researchers from the amazing New York University. The goal was to show a neural network video footage of lots and lots of fluid and smoke simulations and have it learn how the dynamics work to the point that it can continue and guess how the behavior of a smoke puff would change in time. We stop the video and it would learn how to continue it, if you will. Now that is a tall order if I've ever seen one. Most of this episode will not be about the technical details of this method, but about the importance and ramifications of such a technique. And since almost all the time our episodes are about already published works, it also makes a great case study on how to evaluate and think about the merits and shortcomings of a research project that is still in the works. This definitely is an interesting take as normally we use neural networks to solve problems that are otherwise close to impossible to tackle. Here the neural networks are applied to solve something that we already know how to solve, and the question immediately comes to mind why would anyone bother to do that? We've had the very least 20 episodes on different kinds of incredible fluid simulation techniques, so it is abundantly clear that this is a problem that we can solve. However, the neural network does not only solve it correctly in a sense that the results are easily confused with real footage, but what's more, the execution time of the algorithm is in the order of a few milliseconds. For a reasonably sized simulation, this normally takes several minutes. With traditional techniques it does something that we already know quite well how to do, but it does it better in many regards. Loving the idea behind this work Training is a pre processing step that is a long and arduous process that only has to be done once and afterwards querying the neural network that is predicting what happens next in the simulation runs almost immediately. In any case in way less time than calculating all the forces and pressures in the simulation while retaining high quality results. It is like the preparation for an exam that may take weeks, but but when we are finally there in the examination room, if we are well prepared, we make short work of the puny questions the professor has presented us with. I am quietly noting that during my college years I was also studying the beautiful Navier Stokes equations and even as a highly motivated student it took several months to understand the theory and write my first fluid simulator. This neural network can learn something very similar in a matter of days. What a stunning and may I say humiliating revelation. Note that this piece of work has not yet been peer reviewed. There are some side by side comparisons with real simulations to validate the accuracy of the algorithm, but more rigorous analysis is required before publishing. The failure cases for classical handcrafted techniques are easier to identify because of the fact that their mathematical description is available for scrutiny. In the case of a neural network, this piece of mathematics is also there, but it's not intuitive for human beings. Therefore it is harder to assess when it works well and when it is expected to break down. We should be particularly vigilant about this fact when evaluating a task performed by any kind of neural network based learning algorithm. For now, the results look quite reassuring. Even the phenomenon of a smoke puff bouncing back from an object is modeled with high fidelity. There was a loosely related work from the ETH Zurich and Disney Research in Switzerland, and enumerating the differences is a bit too technical for such a short video, but I have included it in the video description box for the more curious fellow scholars out there. Now. You might have noticed the lack of the usual disclaimer in the thumbnail image stating that I did not take any part in the project, which was not the case this time. I feel that it is important to mention my affiliation, even though my role in this project has been extremely tiny. You can read about this in the acknowledgment section of the paper. Needless to say, all the credit goes to the authors of this paper for this amazing idea. I envision all kinds of interactive digital media, including the video games of the future, being infused with such neural networks for real time fluids and smoke simulations. And let's not forget that this is only the first step. We haven't even talked about other kinds of perhaps learnable physical simulations with collision detection, shattering glassy objects and gooey soft body simulations. And we also have seen the very first results with light simulation pipelines that are augmented with neural networks. I think it is now a thinly veiled fact that I am extremely excited for this and this piece of work is not a destination but a stepping stone towards something truly remarkable. Thanks for watching and for your generous support and I'll see you next time. Sa.
