A Neural Ordinary Differential Equations Approach for 2D Flow Properties Analysis of Hydraulic Structures

In hydraulic engineering, the design and optimization of weir structures play a critical role in the management of river systems. Weirs must efficiently manage high flow rates while maintaining low overfall heights and predictable flow behavior. Determining upstream flow depths and discharge coefficients requires costly and time-consuming physical experiments or numerical simulations. Neural Ordinary Differential Equations (NODE) can be capable of predicting these flow features and reducing the effort of generating experimental and numerical data. We propose a simulation based 2D dataset of flow properties upstream of weir structures called FlowProp. In a second step we use a NODE-based approach to analyze flow behavior as well as discharge coefficients for various geometries. In the evaluation process, it is evident that the aforementioned approach is effective in describing the headwater, overfall height and tailwater. The approach is further capable of predicting the flow behavior of geometries beyond the training data. Project page and code: https://github.com/SEilermann/FlowProp