Uncovering the mechanisms of cross-shore sediment transport through interpretable machine learning with limited laboratory data

Cross-shore sediment transport, a key driver of coastal morphodynamic evolution, remains challenging to model due to the simultaneous contributions of multiple factors, which collectively result in a highly nonlinear dynamic. While previous studies have successfully estimated net cross-shore transport rates, their applicability remains limited to specific hydrodynamic and sediment conditions. This work explores the use of interpretable machine learning (ML) methods for this purpose. To this end, we evaluate six models with varying capacity using the SANTOSS dataset. LightGBM and Gaussian Process Regression achieve the best performance, owing to their ability to capture nonlinear processes. The explainability of LightGBM is analyzed using game theory-based Shapley values for both overall and interaction effects of factors across five distinct scenarios. Overall Shapley values quantify individual factor contributions, while interaction values reveal pairwise synergies. The model consistently identifies the dominant factors under diverse conditions and assigns low importance to irrelevant ones. The interaction Shapley values accurately capture key factors’ couplings, demonstrating the potential of interpretable ML to uncover underlying physical mechanisms of cross-shore sediment transport even with limited data. Thus, ML serves not only as a predictive tool but also as a method to explore the processes governing cross-shore sediment transport.