DDPS | Differentiable Physics Simulations for Deep Learning
Abstract from Speaker:
In this talk I will focus on the possibilities that arise from recent advances in the area of deep learning for physical simulations. In this context, especially the Navier-Stokes equations represent an interesting and challenging advection-diffusion PDE that poses a variety of challenges for deep learning methods.
In particular, I will focus on differentiable physics solvers within the larger field of differentiable programming. Differentiable solvers are very powerful tools to guide deep learning processes, and support finding desirable solutions. The existing numerical methods for efficient solvers can be leveraged within learning tasks to provide crucial information in the form of reliable gradients to update the weights of a neural networks. Interestingly, it turns out to be beneficial to combine supervised and physics-based approaches. The former poses a much simpler learning task by providing explicit reference data that is typically pre-computed. Physics-based learning on the other hand can provide gradients for a larger space of states that are only encountered during training runs. Here, differentiable solvers are particularly powerful to, e.g., provide neural networks with feedback about how inferred solutions influence the long-term behavior of a physical model.
I will demonstrate this concept with several examples from learning to reduce numerical errors, over long-term planning and control, to generalization. I will conclude by discussing current limitations and by giving an outlook about promising future directions.
Speaker Bio:
Nils Thuerey is Associate-Professor at the Technical University of Munich (TUM). He focuses on deep-learning methods for physical systems, with an emphasis on fluid flow problems. Beyond latent-space simulation algorithms and generative models, he's currently especially interested in learning algorithms powered by differentiable solvers.
LLNL-VIDEO-819822
Видео DDPS | Differentiable Physics Simulations for Deep Learning канала Livermore Lab Events
In this talk I will focus on the possibilities that arise from recent advances in the area of deep learning for physical simulations. In this context, especially the Navier-Stokes equations represent an interesting and challenging advection-diffusion PDE that poses a variety of challenges for deep learning methods.
In particular, I will focus on differentiable physics solvers within the larger field of differentiable programming. Differentiable solvers are very powerful tools to guide deep learning processes, and support finding desirable solutions. The existing numerical methods for efficient solvers can be leveraged within learning tasks to provide crucial information in the form of reliable gradients to update the weights of a neural networks. Interestingly, it turns out to be beneficial to combine supervised and physics-based approaches. The former poses a much simpler learning task by providing explicit reference data that is typically pre-computed. Physics-based learning on the other hand can provide gradients for a larger space of states that are only encountered during training runs. Here, differentiable solvers are particularly powerful to, e.g., provide neural networks with feedback about how inferred solutions influence the long-term behavior of a physical model.
I will demonstrate this concept with several examples from learning to reduce numerical errors, over long-term planning and control, to generalization. I will conclude by discussing current limitations and by giving an outlook about promising future directions.
Speaker Bio:
Nils Thuerey is Associate-Professor at the Technical University of Munich (TUM). He focuses on deep-learning methods for physical systems, with an emphasis on fluid flow problems. Beyond latent-space simulation algorithms and generative models, he's currently especially interested in learning algorithms powered by differentiable solvers.
LLNL-VIDEO-819822
Видео DDPS | Differentiable Physics Simulations for Deep Learning канала Livermore Lab Events
Показать
Комментарии отсутствуют
Информация о видео
Другие видео канала
"From Berkeley to Berlin" Part 6 — Tom RamosHEDS | Relativistic plasma mirrors for high-power ultrashort pulses from UV to soft x-rayDDPS | Model reduction of partial differential equations via optimization-based feature trackingDDPS | Cheap and robust adaptive reduced order models for nonlinear inversion and designHEDS | Stellar-Relevant Emission-Based Opacity Experiments at the Orion Laser FacilityCGSR | China’s Emergence as a Second Nuclear Peer: Implications for US Nuclear DeterrenceHEDS | Quantitative Studies of Supersonic Microparticle Impacts on MetalsCGSR | Cyberspace Escalation: Ladders or Lattices? Feat. Olesya Tkacheva & Martin LibickiCGSR | Hinge Points: An Inside Look at North Korea’s Nuclear ProgramMFEM Workshop 2022 | Shape and Topology Optimization Powered by MFEMHEDS | Exploring the most Extreme Conditions of Matter with ultra-bright X-rays2022 Bay Area Research SLAM!2022 DSI Summer Slam | Marina DunnHEDS | Stochastic sampling of dense matterCGSR | Arctic: Toward the End of the Exception Strategic, Nuclear and Maritime Issues in the RegionHEDS | An alternate approach to the ignition regime for inertial confinement fusioNCI Seminar: 3D Imaging of Fruit Fly Neurons and MusclesWiDS Livermore 2023 | Datathon Speaker: Dr. Suzanne SindiAPAC Fireside Chat: Tammy MaLLESA Author Series | Chris ColbyDDPS | Neural Galerkin schemes with active learning for high-dimensional evolution equations