Hybrid integration of Computational Fluid Dynamics and Artificial Neural Networks for enhanced predictive modelling in hydraulic engineering

In recent years, the accurate prediction of fluid flow and the hydraulic properties of flow have become increasingly important for the management of underground and surface water systems and hydraulic structures. Computational Fluid Dynamics (CFD) has proven to be one of the most influential simulation techniques in the analysis of complex flow interactions in canals, spillways, reservoirs and other hydraulic systems. Despite its proven reliability and accuracy, CFD is still computationally intensive, particularly for real-time monitoring or large-scale water collection and storage systems. To overcome these limitations, researchers have explored the potential of Artificial Neural Networks (ANNs), which subdivide the ability of this method to approximate nonlinear functions based on input-output relationships. The integration of CFD with ANN represents a promising solution that combines the physical reliability of numerical models with the speed and adaptability of data-driven approaches. The increasing number of complex flow solutions such as air-water interphase, air entrapment, bubble formation in a high-velocity flow, hydraulic jump, turbulent flow and cavitation problems, especially in flood events, require advanced simulation approaches such as CFD to solve the problem from the fundamental aspects to the application for engineering decision-making processes (Chanson, 2022; Chanson et al., 2021; Chanson & Shi, 2022). This selection was made due to the excellent simulation capabilities, reliability and realistic solutions of CFD for fluid dynamic problems (Mozaffari et al., 2022; Sharifi, 2025). However, its application is often limited by high computational costs and the requirement for expertise in setting up the system and validating the model (collected data from the physical inspection on site or the constructed, scaled-down physical model). This therefore represents a significant obstacle to the development of early warning systems, as this model offers a time-dependent approach. These limitations are particularly restricted to real-time predictions or decision support systems for managing hydraulic infrastructures. Conversely, ANN can provide fast approximations of system behaviour if the system is properly trained and sufficient data is available (Jabbari and Bae, 2018; Spiridonov et al. 2020). However, individual ANN models often have problems with reliability and accuracy when exposed to complex flow scenarios without physical laws embedded in the numerical models (Frnda et al., 2022). An urgent challenge is therefore to develop hybrid methods that integrate CFD with ANNs to capitalise on the strengths of both data-driven and physics-based approaches.