|Time:||June 22, 2022, 4:00 p.m. (CEST)|
|Download as iCal:||
While there's growing interest in using machine learning to speed up numerical simulations, many existing methods are only accurate in a very narrow range around the training data. In practice, this often means running large numbers of expensive reference simulations to generate targeted training data for each problem, negating performance benefits of using learned methods. But it doesn't have to be that way. In this talk, we'll explore how to construct GNN-based models which learn reusable knowledge and generalize well; and we'll see how exploiting a set of very general inductive biases allows us to accurately predict dynamics of significantly bigger and more complex systems than seen in training. I will show some demonstrations on a range of physical systems, from aerodynamics over structural mechanics to cloth, and discuss what we can do with these models beyond forward prediction.
Tobias Pfaff is a senior research scientist at DeepMind, working on RL and structured models, in particular GNN physics models. His background is in physics simulation; in his PhD (ETH Zurich) and postdoc (UC Berkeley), he studied turbulent fluids, fracture and cloth simulation for computer graphics, and co-authored the simulation frameworks Mantaflow (the default fluid solver in Blender) and ArcSim.