Machine learning algorithms in sediment transport modeling: a review

The purpose of this review is to clarify how emerging hybrid, physics-informed, and network-aware machine learning (ML) approaches are reshaping sediment-transport modeling. The central research question is: How can ML methods, when aligned with data regimes, physical constraints, and spatial structure, improve sediment-transport prediction across diverse river environments? To address this, the review reorganizes existing studies into six environmentally grounded clusters and conducts a cross-study comparison to identify model performance patterns, limitations, and future directions. A survey of peer-reviewed literature was conducted, covering ML applications in sediment-transport modeling across rivers, reservoirs, estuaries, and coastal systems. Ensemble models and sequence-based architectures consistently improve performance in event-driven, high-sediment regimes, while non-parametric and time-series models better capture temperature- and melt-controlled dynamics in alpine basins. The cross-study comparison highlights strong performance of ensemble and hybrid methods, yet also reveals variability due to inconsistent validation strategies and data heterogeneity. Because the underlying studies differ in targets, input features, and validation procedures, this comparison is used to identify broad performance patterns rather than to rank individual algorithms, and the associated caveats are discussed. This review shows that hybrid, physics-informed, and network-aware approaches can deliver more robust and transferable predictions than traditional methods.