Forecasting and Interpolation for Learning Physical Simulation over Meshes

Speaker: Xiao Luo, Ph.D.
IDRE Fellow
Department of Computer Science
University of California Los Angeles
Abstract: This talk discusses the problem of learning-based physical simulation, a crucial task with applications in fluid mechanics and aerodynamics. Recent works typically utilize graph neural networks (GNNs) to produce next-time states on irregular meshes by modeling interacting dynamics and then adopting iterative rollouts for the whole trajectories. Our work proposes a simple yet effective approach named FAIR for long-term mesh-based simulations. Our model employs a continuous graph ODE model that incorporates past states into the evolution of interacting node representations, capable of learning coarse long-term trajectories under a multi-task learning framework. Then, we leverage a channel aggregation strategy to summarize the trajectories for refined short-term predictions, which can be illustrated using an interpolation process. Our method can generate accurate long-term trajectories through pyramid-like alternative propagation between the foresight step and refinement step. Finally, we show the experiments on several benchmark datasets to validate the effectiveness of our method.