AI-enhanced simulation of sediment transport in cold regions

Climate change is rapidly reshaping hydro-geomorphological processes in cold regions. Melting glaciers and thawing permafrost are altering how and when sediment is mobilized, creating sediment supplies that are highly sensitive to warming and shifting precipitation patterns. During heavy rainfall and/or intense melting, this abundant and readily mobilized sediment can lead to substantial increases in sediment fluxes, triggering episodic sediment transport events widely observed in permafrost watersheds. These events are typically characterized by the complex co-occurrence of multiple factors such as transient and complex flow conditions, temporarily enhanced erosivity, and dynamic sediment availability. However, widely applied empirical, process-based, and data-driven sediment-transport models (e.g., rating curves, SAT, HydroTrend, SWAT, WBMsed) commonly assume stationary parameters or simplified process dynamics and tend to underestimate both the magnitude of episodic sediment transport. Artificial intelligence (AI)–based data-driven models, including machine learning and deep learning algorithms, have emerged as powerful tools for suspended sediment concentration modeling due to their ability to represent nonlinear and nonstationary processes. Using twenty years of hydrological observations, we found that the drivers of sediment transport now show distinct seasonal variations. To better capture these complex and seasonal shifting processes, we developed a modified deep learning model to learn seasonal differences in sediment transport and dynamically adjusts its predictive weights. It performs substantially better than current widely applied models including rating-curves, processes-based and random forest models, particularly during extreme sediment transport. Our results demonstrate the promise of integrating AI with process understanding to simulate highly variable sediment dynamics under changing climate and cryosphere conditions.