Yuan et al. (2026) A dynamic-gated Mixture-of-Experts framework improves and interprets daily streamflow simulation

Identification

• Journal: Communications Earth & Environment
• Year: 2026
• Date: 2026-07-09
• Authors: Wenrui Yuan, Shi Hu, Chesheng Zhan, Zhonghui Lin
• DOI: 10.1038/s43247-026-03799-z

Research Groups

• Department of Hydrology, University of California
• National Center for Atmospheric Research (NCAR)
• US Geological Survey (USGS)

Short Summary

This study presents HydroMoE, a dynamic-gated Mixture-of-Experts framework that improves and interprets daily streamflow simulation. The framework pairs process-based and neural experts for each hydrological subprocess, allowing the model to balance physical and data-driven representations.

Objective

• Research question: How can we develop a hybrid modeling approach that combines the strengths of process-based models (PBMs) and deep learning models to improve daily streamflow prediction?
• Principal hypothesis: A dynamic-gated Mixture-of-Experts framework can adaptively allocate weights between physical and neural experts, enabling more accurate and physically interpretable predictions.

Study Configuration

• Spatial Scale: Continental United States
• Temporal Scale: Daily

Methodology and Data

• Models used: HydroMoE (dynamic-gated Mixture-of-Experts), PBM (process-based model)
• Data sources: CAMELS-US dataset, satellite observations, reanalysis data

Main Results

• Key findings: HydroMoE improves daily streamflow prediction accuracy compared to the process-based baseline and single-module neural ablations.
• Quantitative results: Median NSE of 0.663 and median KGE of 0.638 for HydroMoE, compared to 0.127 and 0.210 for the PBM.

Contributions

• Original value of the article: The study presents a novel hybrid modeling approach that combines process-based and neural representations, enabling more accurate and physically interpretable predictions.
• Contribution to existing literature: The framework addresses the limitations of earlier sequential methods by actively learning hydrological logic and providing insights into underlying physical mechanisms.

Funding

• National Science Foundation (NSF)
• US Geological Survey (USGS)
• Department of Energy (DOE)

Citation

@article{Yuan2026dynamicgated,
  author = {Yuan, Wenrui and Hu, Shi and Zhan, Chesheng and Lin, Zhonghui},
  title = {A dynamic-gated Mixture-of-Experts framework improves and interprets daily streamflow simulation},
  journal = {Communications Earth & Environment},
  year = {2026},
  doi = {10.1038/s43247-026-03799-z},
  url = {https://doi.org/10.1038/s43247-026-03799-z}
}