Deep Learning for Fluid dynamics simulation
Steeven Janny
Title: Deep Learning for Fluid dynamics simulation
Abstract: Fluid dynamics simulation traditionally relies on computationally intensive numerical models solving the Navier-Stokes equations, posing challenges in complexity and time consumption. Recent advancements integrate machine learning, notably graph neural networks (GNNs) and attention mechanisms, to address these challenges. Yet, the domain is still in its infancy. During my Ph.D., we contributed to this field by proposing new models and datasets focused on simulating fluid dynamics. We introduce EAGLE, a large-scale dataset comprising 1.1 million 2D meshes simulating unsteady fluid dynamics interacting with nonlinear scene structures. We propose a novel mesh transformer that leverages node clustering, graph pooling, and global attention to learn long-range dependencies efficiently. Additionally, we address the limitations of fixed mesh computations in fluid dynamics simulations by proposing a data-driven approach. We formulate the task as a double observation problem and devise a solution with interlinked dynamical systems operating on sparse positions and continuous domains. This enables predictions in a continuous spatial and temporal domain, trained on sparse observations, and capable of forecasting and interpolating solutions from initial conditions.
Website: https://steevenjanny.github.io/
In Room M7.101 of Monod campus, ENSL.