Dr. Károly Zsolnai-Fehér explores whether DeepMind's AI can learn physics by running neural networks on fluid simulations. The research shows AI can perform complex simulations faster than traditional methods, potentially transforming how scientists model physical systems.
Dear fellow scholars, this is Two Minute Papers with Dr. Zsolnai. If you have been watching this series for a while, you know very well that I love learning algorithms and fluid simulations. But do you know what I like even better? Learning algorithms applied to fluid simulations. So I couldn't be happier with today's paper. We can create wondrous fluid simulations like the ones you see here by studying the laws of fluid motion from physics and writing a computer program that contains these laws. However, I just mentioned learning algorithms. How do these even come to the picture? If we can write a program that simulates the laws, why would we need learning based algorithms? This doesn't seem to make any sense. You see, in this task, neural networks are applied to solve something that we already know how to solve. However, if we use a neural network to perform this task, we have to train it, which is a long and arduous process. I hope to have convinced you that this is a bad, bad idea. Why would anyone bother to do that? Does this make any sense? Well, it does make a lot of sense, and the reason for that is that this training step only has to be done once. And afterwards, querying the neural network that is predicting what happens next in the simulation runs almost immediately. This takes way less time than calculating all the forces and pressures in the simulation while retaining high quality results. This earlier work from last year absolutely nailed this problem. Look, this is a scene with the boxes it has been trained on. And now let's ask it to try to simulate the evolution of significantly different shapes. Wow. It not only does well with these previously unseen shapes, but it also handles their interactions really well. But there was more. We could also train it on a tiny domain with only a few particles, and then it was able to learn general concepts that we can reuse to simulate a much bigger domain and also with more particles. Fantastic. This was a simple general model that truly is a force to be reckoned with. Now this is a great leap in neural network based physics simulations, but of course, not everything was perfect there. For instance, over longer time frames, solids became incorrectly deformed. And now a newer iteration of a similar system just came out from DeepMind's research lab that promises to extend these neural networks for an incredible set of use aerodynamics, structural mechanics, class simulations, and more. I am very excited to see how far they have come since. So let's see how well it does first with rollouts, then with generalization experiments. Here is the first rollout experiment. So what does that mean? And what are we seeing here on the left, you See a verified handcrafted algorithm performing the simulation. We will accept this as the true data. And on the right, the AI is trying to continue the initial simulation. But there is one problem, and that problem is that the AI was only trained on short simulations with 400 time steps. That's only a few seconds and unfortunately this test will be a hundred times longer. So it only learned on short simulations. Can it manage to run a longer one and remain stable? Well, that will be tough, but so far so good. Still running. My goodness, this is really something. Still running. And it's very close to the ground truth. Ok, that is fantastic. But that was just a piece of cloth. What about interactions with other objects? Well, let's see. I'll stop the process here and there so we can inspect the differences. Again, flying colors. Loving it. And apparently the same can be said for simulations in structural mechanics and incompressible fluid dynamics. Now there is one more important lesson here. To be able to solve such a wide variety of simulation problems, we need a bunch of different handcrafted algorithms that took many, many years to develop. But this one neural network can learn and perform them all. And it can do it ten to a hundred times quicker. And now comes the second half, generalization experiments. This means a simulation scenario which shapes that the algorithm has never seen before. And let's see if it obtained general knowledge of the underlying laws of physics to be able to pull this off. Oh my, look at that. Even the tiny piece that is hanging off of the flag is simulated nearly perfectly in this one. They gave it different wind speeds and directions that it hadn't seen before. And not only that, but we are varying these parameters in time and it doesn't even break a sweat. And hold on to your papers because here comes my favorite. It can even learn on a small scale simulation with a simple rectangular flag. And now we throw at it a much more detailed cylindrical flag with tassels. Surely this will be way beyond what any learning algorithm can do today. And ok, come on. I am truly out of words. Look, so now this is official. We can ask an AI to perform something that we already know how to do. And it will not only be able to reproduce similar simulations, but we can even ask things that were previously quite unreasonably outside of what it had seen. And it handles all these with flying colors. And it does this much better than previous techniques were able to. And it can learn from multiple different algorithms at the same time. Wow, what a time to be alive. This video has been supported by weights and biases. Check out their recent offering, Fully Connected, a place where they bring machine learning practitioners together to share and discuss their ideas, learn from industry leaders, and even collaborate on projects together. You see, I get messages from you fellow scholars telling me that you have been inspired by the series but don't really know where to start. And here it is. Fully Connected is a great way to learn about the fundamentals, how to reproduce experiments, get your papers accepted to a conference and more. Make sure to visit them through WNB Me papers or just click the link in the video description. Our thanks to Weights and Biases for their long standing support and for helping us make better videos for you. Thanks for watching and for your generous support and I'll see you next time.
