Conclusion

This study reconstructed a chemically and geographically harmonized national dataset of 5085 paired sediment–water records and developed a multi-branch multi-head attention (MB-MHA) framework to predict sediment–water partitioning (_K_d) of ECs in natural river systems. The MB-MHA model improved predictive accuracy and clarified how microscale molecular attributes, mesoscale sediment–water features, and macroscale basin environments jointly influence partitioning behavior. Integration with molecular dynamics simulations further revealed class-dependent differences in dominant controlling factors governing sediment–water partitioning, ranging from primarily molecular descriptor-driven behavior in ABs to ion- and structure-dependent interactions in PFASs, and more multifactor-coupled influences in EDCs. These differences are consistent with the observed variation in spatial heterogeneity across EC classes at the basin scale. By linking molecular-level interpretations with basin-scale predictive mapping, the framework identified hotspots with sediment-accumulation potential (high _K_d) and enhanced aqueous mobility (low _K_d), offering risk-relevant insights for EC monitoring and management. These results provide information relevant to EC monitoring and management. Future work should incorporate mixture effects and dynamic contaminant-transformation processes to extend the framework to complex aquatic systems.