Figure showing main results from: Chuang, P.Y. and Barba, L.A., 2023. Predictive Limitations of Physics-Informed Neural Networks in Vortex Shedding. arXiv preprint arXiv:2306.00230. —Snapshots of flow fields (u velocity) for flow around a circular cylinder from PetIBM, unsteady PINN, and data-driven PINN. At t = 10 and t = 50, the unsteady PINN matches the PetIBM simulation, indicating that it can solve unsteady equations. However, the data-driven PINN does not produce meaningful results at these time points, suggesting its inability to extrapolate backward in time. At t = 140, vortex shedding occurs, but the unsteady PINN cannot capture it, while the data-driven PINN qualitatively displays shedding. At t = 190, the behavior of data-driven PINN tends towards that of the unsteady PINN, which behaves like a steady-state solver.
Group website: http://lorenabarba.com
Group website: http://lorenabarba.com