BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:CESS Seminar - Learned Models for Physical Simulation and Design
DTSTART:20220513T121500
DTEND:20220513T130000
DTSTAMP:20260407T091258Z
UID:4712cd5fa731c7700b51e0f1f821da629d04ea23ef716e0c6cf4b3ad
CATEGORIES:Conferences - Seminars
DESCRIPTION:Dr Kimberly Stachenfeld\, Research Scientist at DeepMind\n\n\n
 \nAbstract\n\n\n\nSimulation is important for countless applications in sc
 ience and engineering\, and there has been increasing interest in using ma
 chine learning to produce learned simulators to produce simulations more e
 fficiently than classical simulators\, distill dynamics into a differentia
 ble model\, or learn simulators from real world data. In the first part of
  our talk\, I will describe our recent work on training learned models for
  efficient turbulence simulation. Turbulent fluid dynamics are chaotic an
 d therefore hard to predict\, and classical simulators typically require 
 expertise to produce and take a long time to run. We found that learned CN
 N-based simulators can learn to efficiently capture diverse types of turbu
 lent dynamics at low resolutions\, and that they capture the dynamics of a
  high-resolution classical solver more accurately than a classical solver 
 run at the same low resolution. We also provide recommendations for produc
 ing stable rollouts in learned models\, and improving generalization to ou
 t-of-distribution states. In the second part of the talk\, I will discuss 
 our recent work using learned simulators for inverse design. In this work\
 , we combine Graph Neural Network (GNN) learned simulators [Sanchez-Gonzal
 ez et al 2020\, Pfaff et al 2021] with gradient-based optimization in orde
 r to optimize designs in a variety of complex physics tasks. These include
  challenges designing objects in 2D and 3D to direct fluids in complex w
 ays\, as well as optimizing the shape of an airfoil. We find that the lear
 ned model can support design optimization across 100s of timesteps\, and t
 hat the learned models can in some cases permit designs that lead to dynam
 ics apparently quite different from the training data.\n\n\n\n \n\n\n\nBi
 ography\nKimberly Stachenfeld is a Research Scientist at DeepMind. Her res
 earch focuses on building structured deep learning models of complex syste
 ms. She is particularly interested in learned simulation and graph neural 
 networks. She also works on problems in Computational Neuroscience\, where
  she models graph-based learning and reasoning in humans and animals. She 
 received her Ph.D. in Computational Neuroscience from Princeton Universit
 y in 2018\, and her B.A./B.S. in Mathematics/Chemical and Biological Engi
 neering from Tufts University in 2013.\n\n\n
LOCATION:GC B3 30 https://plan.epfl.ch/?room==GC%20B3%2030
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR