Is Simulating A Jelly Sandwich Possible? 🦑

Dear fellow scholars, this is Two Minute Papers with Dr. Zsolnai. Today we are going to torment a virtual armadillo. Become happy or sad, depending on which way we are bending, create a ton of jelly sandwiches, design a crazy bridge and more. So what is going on here? Well, this new paper helps us enhance our physics simulation programs by improving how we evaluate derivatives. Derivatives describe how things change. Now, of course, evaluating derivatives is not new, it's not even old. It is ancient. Here you see an ancient technique for this and it works well most of the time, but. Whoa. Well, this simulation blew up in our face. So yes, it may work well most of the time, but unfortunately not here. Now, if you're wondering what should be happening in this scene, here is the reference simulation that showcases what should have happened. Yes, this is the famous pastime of the computer graphics researcher tormenting virtual objects basically all day long. So I hope you are enjoying this reference simulation because this is a great reference simulation. It is as reference as a simulation can get. Now, I hope you know what's coming. Hold on to your papers because this is not the reference simulation. What you see here is the new technique described in this paper. Absolutely amazing, actually. Let's compare the two. This is the reference simulation for real this time. And this is the new complex step finite difference method. The two are so close, they are essentially the same. I love it. So good. Now, if this comparison made you hungry, of course we can proceed to the jelly sandwiches. Here is the same scene simulated with a bunch of previous techniques. And my goodness, all of them look different. So which of these jelly sandwiches is better? Well, the new technique is better because this is the only one that preserves volume properly. This is the one that gets us the most jelly. With each of the other methods, the jelly either reacts incorrectly or at the end of the simulation, there is less jelly than the amount we started with. Now, you're probably asking, is this really possible? And the answer is yes, yes, it is. What's more, this is not only possible, but this is a widespread problem in physics simulation. Our seasoned fellow scholars had seen this problem in many previous episodes. For instance, here is one with the tragedy of the disappearing bunnies. And preserving the volume of the simulated materials is not only useful for jelly sandwiches, but it is also useful for doing extreme yoga. Look, here are a bunch of previous techniques trying to simulate this. And what do we see? Extreme bending. Extreme bending. And even more extreme bending. Good, I guess. Well, not quite. This yoga shouldn't be nearly as extreme as we see here. The new technique reveals that this kind of bending shouldn't happen in given these material properties. And wait, here comes one of my favorites. The new technique can also deal with this crazy example. Look, a nice little virtual hyper elastic material where the bending energy changes depending on the bending orientation, revealing a secret or two secrets. As you see, it does not like bending to the right so much, but bending to the left. Now we're talking. And it can also help us perform inverse design problems. For instance, here we have a hyper elastic bridge built from over 20,000 parts. And here we can design what vibration frequencies should be present when the wind blows at our bridge. And here comes the coolest part. We can choose this in advance. And then the new technique quickly finds the suitable geometry that will showcase the prescribed vibration types. And it pretty much converges after four to six Newton iterations. What does that mean? Yes, it means that the technique comes up with an initial guess and it needs to refine it only four to six times until it comes up with an excellent solution. So better hyper elastic simulations quickly and conveniently. Yes, please sign me up right now. And what would you use this for? Let me know in the comments below. 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, 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'll see you next time.