DDPS | Learning to accelerate large-scale physical simulations in fluid and plasma physics

Description: Simulating the time evolution of large-scale physical systems is crucial in many scientific and engineering domains, such as in fluid dynamics and plasma physics. Typically, domain-specific classical numerical solvers are employed to simulate such systems, which may require massive computational resources to simulate. Recently, deep learning surrogate modeling has emerged as a useful tool to complement or speed up the simulations, but scaling it to simulate larger systems still faces several challenges: how to learn faithful long-term evolution, and how to model multi-scale dynamics. In this talk, I will present our works in a subsurface fluid system and a plasma system that aim to address these challenges. In a subsurface fluid system that consists of millions of cells per time step, we introduce a hybrid graph network simulator (HGNS), which consists of a graph neural network (GNN) for fluid dynamics and a 3D-U-Net for pressure dynamics, to model the complex and heterogeneous structure and dynamics in the subsurface. We introduce a multi-step based objective to improve its long-term rollout, and sector-based training to make training such system possible in GPU. Experiments show that our method is able to scale up to millions of cells per time step, outperforms strong baselines of learning-based methods in long-term rollout, and achieves up to 18x speedup compared to the classical solvers. In a laser-plasma interaction system whose dynamics is multiscale, we introduce a hybrid particle + fluid representation, which uses the Particle-in-Cell solver to model kinetics of the few but highly energetic particles, and convolutional neural networks to model the evolution of the moments of the dominant thermal particles in each cell, together with their injection from fluid state to particle state. We also demonstrate preliminary results for this system.

Bio: Tailin Wu is a postdoctoral scholar in the Department of Computer Science at Stanford University, working with professor Jure Leskovec. His research interests include representation learning and machine learning for scientific simulations and inverse design, using tools of graph neural networks and information theory. He obtained his PhD in MIT Physics, where his thesis focused on application of machine learning to physics and introducing physics insights/techniques into machine learning.

LLNL-VIDEO-836058

#LLNL #PhysicalSimulations #PlasmaPhysics

[divyanshuvyas]

Hi, great session! Thanks. I have a question regarding the Subsurface Simulations using HGNS part. What do you mean by Global dynamics vs Local Dynamics? Why GNN is preferred for local dynamics as compared to 3D Unet for Pressure evolution i.e global dynamics? Am I missing something here.
Thanks in advance.