SedimentNet — a 1D-CNN machine learning model for prediction of hydrodynamic forces in rapidly varied flows

In natural free surface flows, sediment particles in the surface layer of a sediment bed are moved and entrained by the fluctuating hydrodynamic forces, such as lift and drag, exerted by the overlying flow. Accurate prediction of near-bed hydrodynamic forces in rapidly varied flows is vital for coastal sediment transport and morphodynamics. Directly measured hydrodynamic forces within the rapidly varied flows over rough bed layer have been limited by previous spatial averaging shear force studies. Therefore, the direct measurements were designed and adapted to estimate tidal bore forces, including longitudinal (drag) and vertical (lift) force on near-bed sediment particle. Specially designed experiments were conducted to measure the instantaneous forces using a highly sensitive force sensor assembled with a target sphere. A novel 1D-CNN model (i.e. SedimentNet) has been developed for the prediction of hydrodynamic forces and compared with existing machine learning models (i.e. Decision Tree (DT), Random Forest (RF), Multilayer Perceptron (MLP), Support Vector Regressor (SVR), XGBoost, \(k\)-Nearest Neighbour (\(k\)-NN)). The parameters affecting near-bed hydrodynamic forces are also analysed. In the context of machine learning, both conventional dataset split and fivefold cross-validation approaches were implemented. The results indicated that the proposed SedimentNet was able to achieve marginally better cross-validation performance (i.e. \({R}^{2}\) score of 0.77 for drag force, \({R}^{2}\) score of 0.96 for lift force) for the prediction of drag and lift forces. RF and XGBoost were the second best models with \({R}^{2}\) score of 0.73 and 0.95 for drag and lift force prediction, respectively. Results also showed the potential of machine learning models for the efficient prediction of complex hydrodynamic forces in a coastal environment. A use-case edge computing solution for the reported machine learning-based prediction of hydrodynamic forces has also been proposed and discussed to demonstrate the practical implementability of the presented research.