Welcome everybody to another episode of Beyond the Patterns.

Today we have a really exciting guest.

It's going to be Niels Thürer from TU Munich and he is working on differentiable physical

simulation.

So I'll tell you there's exciting things coming up today.

You'll see plenty of videos and how to integrate a physical solver differentially into a deep

neural network.

Starting work, Nils Thuray is an associate professor at the Technical University of Munich.

He focuses of course on deep learning methods for physical systems with an emphasis on fluid

flow problems.

Beyond latent space, simulation algorithms and generative models, he is currently especially

interested in learning algorithms powered by differentiable solvers.

So today's presentation will therefore also be entitled Differentiable Physics Simulations

for Deep Learning.

Nils, it's a great pleasure to have you here and the stage is yours.

Thank you for the introduction, Nils.

It's great to be virtually in Erlangen at least.

I studied in Erlangen and did my PhD so I have a lot of good memories and I hope to

be back in person at some point but yeah, this is also nonetheless a very good opportunity.

I want to talk about some of the works that we're busy with in my group and not surprisingly

which I think are a very interesting direction at the moment for research and as you can

see from the title that is Coupling Differentiable Physics Simulations with Deep Learning Algorithms.

Yeah, let me start very broadly.

Why are we interested in this?

We typically looking at physical phenomena that are everywhere around us in our everyday

life.

The air around you in the room you're in that's going into your lungs, you typically don't

see it but if you have something like smoke, you see that typically they're very complicated

motions beneath it.

It's very important for classical questions such as how much lift do we need to keep an

airplane up in the air and interestingly also actually for the glass of water on your desk,

it obeys very similar descriptions typically for liquids, we are in the air, we're looking

at the liquid so it's a two-phase problem, we're looking at the interface but nonetheless

the equations are the underlying physical models are very similar.

In robotics just to give one additional area, there are also all kinds of interesting problems.

This robotic fish is a very good example but all kinds of robots in the practice have to

interact with the flow be it around an autonomous vehicle or some autonomous creatures in the

water.

The point being there are all kinds of applications here although fluids might sound a bit specialized

at first, it's really a very broad topic from chemical reactions over plasma physics, energy

topics to medical ones, it comes up a lot and the underlying physical model which I

have to show here, this is really the basis of everything I'll be dealing with in the

following are the Navier-Stokes equations.

I don't want to go into any additional detail here, it's just that we have classical equations

here they're very established and we can approximate them in the computer.

So in a way, now the interesting question at the moment is given all these exciting

new deep learning algorithms, how do these go into the picture here?

And I want to take a fairly pragmatic standpoint here so these physical models, they are great,

we want to keep them.

And also we have over the course of many decades here very powerful numerical methods to find