Towards capturing bedform transition: harnessing capabilities of CFD bedform models and machine learning

Our understanding of bedform processes and their associated effect on bedform roughness is limited, and accounts for large uncertainties in hydraulic roughness computation. It is a standard practice in hydraulic modelling to consider hydraulic roughness as a roughness coefficient and to calibrate the model to this coefficient. Such an approach is empirical and does not well capture the physical processes involved in hydraulic roughness dynamics. When bedforms are present, they can account for a significant portion of hydraulic roughness. Consequently, when bedform transitions occur, an abrupt and significant disruption in the hydraulic roughness regime occurs; affecting our water management applications, such as navigability, flood risk management, sediment transport, etc. Bedform transitions are rarely captured, either in laboratory or in real-scale river channels. As such, our understanding of such transition behaviour is further constrained.

In this research, we modelled a CFD physics-based bedform model for the Chiyoda channel, Japan based on previous study of Yamaguchi et al. (2019). The model configuration and results of that study had been validated. The CFD model was initialized at flat-bed condition and run till a dynamic equilibrium in dune regime was obtained. In our research, we captured the bedform in this simulation in each time step, effectively obtaining a timeseries of bed evolution from flatbed regime to dune regime.

It is hoped, the use of physics-based CFD models can simulate the physical processes that invoke bedform transitions. As these have not been easily observed in the field or in lab, the simulations can provide an important insight into these complex processes. This is particularly important in the context of changing hydraulic regimes under the changing climate scenario – possibly making past calibrations of river systems incompatible in the future. An alternative to physics-based (CFD) model is the use of a data-driven model (using machine learning techniques). The use of surrogate machine learning models that capture the behaviour of these physics-based (CFD) models, provides an advantage in terms of computational cost and computational time.

We also developed a proof-of-concept artificial neural network ML models to predict dune height and mean flow depth respectively based on the CFD model results as input. Several models were built using various combinations of input variables: the lagged values of dune height and mean flow depth, mean flow depth or dune height (alternatively), as well as the present and lagged values of spectral power from Fast Fourier Transform spectral analysis. The lagged values of the predicted variable were the most important input parameters compared to other variables. The use of spectral power as predictive variable did not much improve the results, owing to a strong cross-correlation of the parameter with dune height and mean flow depth.

Alternative predictive variables such as stream discharge, Froude number, etc may be considered in future studies to ensure better prediction ability. Validation of these ML and physics-based CFD model results remain a challenge as bedform transition timeseries dataset is not much available. Future outlook of the research in this direction is discussed.