Poudel et al. (2026) An argument for parsimony in differentiable hydrologic models

Identification

• Journal: Hydrology and earth system sciences
• Year: 2026
• Date: 2026-08-14
• Authors: Sandeep Poudel, Scott Steinschneider
• DOI: 10.5194/hess-30-5173-2026

Research Groups

• Department of Civil and Environmental Engineering, Cornell University
• Department of Biological and Environmental Engineering, Cornell University

Short Summary

This study compares static and dynamic differentiable hydrologic models to evaluate their performance in predicting streamflow. The results show that static models can achieve comparable performance to dynamic models while being simpler and more interpretable.

Objective

• Investigate the trade-offs between complexity and performance in differentiable hydrologic modeling.
• Evaluate the role of neural network architecture, process model configuration, and input features on model performance.

Study Configuration

• Spatial Scale: Catchment scale (531 CAMELS-US basins)
• Temporal Scale: Daily to annual time steps

Methodology and Data

• Models used:
  • Multi-Layer Perceptron (MLP) with HBV unit
  • MLP with ensemble of 16 HBV units
  • Long Short-Term Memory (LSTM) with HBV unit
  • LSTM with ensemble of 16 HBV units
• Data sources: CAMELS-US dataset, Daymet meteorological product, USGS streamflow data

Main Results

• Static models can achieve comparable performance to dynamic models in predicting streamflow.
• Ensemble configurations outperform single-unit configurations for both MLP and LSTM architectures.
• The performance advantage of the LSTM over the MLP disappears when using ensemble configurations.

Contributions

• This study provides a systematic comparison of static and dynamic differentiable hydrologic models, highlighting the potential benefits of simpler models while maintaining predictive performance.
• The results suggest that static inputs may function as spatial proxies for generalization in space or as site identifiers when generalizing in time.

Funding

• Not specified

Citation

@article{Poudel2026argument,
  author = {Poudel, Sandeep and Steinschneider, Scott},
  title = {An argument for parsimony in differentiable hydrologic models},
  journal = {Hydrology and earth system sciences},
  year = {2026},
  doi = {10.5194/hess-30-5173-2026},
  url = {https://doi.org/10.5194/hess-30-5173-2026}
}