Evaluation of the Capability of Physics-Aware Neural Networks in Accelerating Flood Simulation Using STE Software

Background and objectives: Floods cause significant damage to urban and rural areas. Rapid analysis and visualization of floodplains during a flood event are essential for identifying threatened areas and assessing potential damages. One of the most critical aspects of rapid flood modeling is the ability to accurately and promptly predict areas at risk. These predictions can warn authorities and residents in vulnerable areas, allowing for timely measures to reduce casualties and financial losses. During a flood, having precise and fast models helps officials make better decisions regarding rescue operations, evacuation, and resource allocation. Extensive research has been conducted on flood modeling, and numerous hydraulic, hydrological, and empirical models have been developed and reviewed. Despite numerous studies, there is still no software or model capable of rapid flood modeling during a flood event. Therefore, this research aims to develop the 2D module of the STE software to achieve accurate and rapid flood modeling using artificial intelligence methods and 2D shallow water equations.
Materials and methods: In this study, the 2D module of the STE software was developed to evaluate the capabilities of two different architectures of Perceptron artificial neural networks, named physics-aware neural networks, in rapid flood modeling in the Eudlo Creek, located in the Sunshine Coast region of Queensland, Australia. The physics-aware neural networks were trained using a genetic algorithm. To test the trained networks and evaluate their ability for rapid flood modeling, their results were compared with those obtained from the finite difference numerical solution. The flood hydrograph used for modeling was obtained from the Australian Bureau of Meteorology for the upstream hydrometric station of the studied river.
Results: The comparison of results obtained from physics-aware neural networks (PANNs) with those from the finite difference method showed that PANNs can reduce the time required for modeling by 50 to 70 percent while maintaining significant accuracy and stability. Increasing the beta parameter in both neural networks enhanced the modeling speed but reduced accuracy. Complex PANNs preserved higher levels of accuracy and stability, especially with larger time steps and higher beta values, resulting in less computational error and outcomes closer to the numerical solution. Increasing the beta parameter significantly increased errors while slightly reducing the modeling completion time. Hence, the optimal beta value for the study area and complex PANNs was determined to be 8. Complex PANNs also demonstrated acceptable accuracy in depicting changes in flow depth and floodplain over time, making them suitable for rapid flood modeling, floodplain mapping, identifying threatened areas, crisis management, and reducing flood damage.
Conclusion: The AI methods examined in this study demonstrated the ability to increase modeling speed and reduce the time required for flood modeling and floodplain mapping while maintaining adequate accuracy. These methods can serve as effective tools for rapid flood modeling, enabling quicker identification of flood-prone areas, timely notifications, and evacuations of at-risk regions, thereby helping to save lives and reduce financial losses.