Physics Informed Neural Networks for Levee Breach Flows

Abstract Submission: Levee breach flooding has a significant impact on the economy and human safety. Levee breach flows are conventionally studied using numerical models that solve shallow water equations and spatially varied flow equation. In this study, Physics-Informed Neural Networks (PINNs) are developed to determine the discharge through levees. The PINN models are developed based on spatially varied flow equation and shallow water equations by assimilating data from a validated 2D numerical model. The PINN model can estimate the water surface profile at the breach section and in the main channel, which aids in understanding the safety of levee structures. Unlike conventional numerical methods, these PINN models do not need to satisfy stability conditions arising from discretization. The PINN models are computationally efficient and can assimilate real-time in-situ observations and remote sensing data. The results of the PINN models, based on spatially varied flow equation and shallow water equations, can help compare accuracy in estimating levee breach flows. Future research will include unsteady flow conditions and their effects on levee breach flows. This expansion can help to further understand the capabilities of PINN models for real-time hydraulic simulation with data assimilation.

Learning Objectives/Expected Outcome (Optional) :