Intelligent prediction of sediment floc settling velocity: Employing explainable machine learning and quantifying the impacts of environmental factors

Settling velocity of flocs resulting from cohesive sediment flocculation, significantly influenced by hydrological factors (e.g., turbulence shear, salinity, suspended sediment concentration and sediment size), is a key parameter to evaluate sediment transport in estuarine and coastal ocean. Measurement and prediction of in-situ settling velocity for sediment flocs are challenging at current stage due to the complex estuarine dynamics and technique constraints. A powerful and interpretable prediction model is urgently required. To meet this requirement, we propose a new paradigm in predicting settling velocity with interpretable machine learning (ML) model by combination of ML with Shapley additive explanation (SHAP) using 4 hydrological parameters as independent input features. Eight ML algorithms (including recently proposed XGBoost, LightGBM and so on) and ensemble learning (Stacking) were employed, with Bayesian Optimization Algorithm to determine the hyper-parameters. XGBoost model was found to have the highest prediction accuracy in terms of 6 performance metrics, among the 8 base learners. Stacking of these models outperform each component model. With the powerful SHAP methods, the proposed paradigm is able to interpret/explain ML model predictions and highlight the highly influential features for the system of sediment dynamics in both local and global points of view. Additionally, the developed ML model provides transparent insights into the contribution of each feature without compromising predictive accuracy. These ML-based models outperform conventional modified Stokes law or empirical equations. Specially, physics-informed neural network (PINN) is employed for prediction of floc settling velocity, which is also a useful approach to overcome the limitations of data driven model characterized by black box. Combinations of ML with physical law and SHAP, the key of the proposed paradigm, could be very powerful tool for researchers on sediment transport.