Feigl et al. (2026) Distilling hydrological and land-surface model parameters from physio-geographical properties using text-generating AI

Identification

• Journal: Nature Water
• Year: 2026
• Date: 2026-02-13
• Authors: Moritz Feigl, Mathew Herrnegger, Karsten Schulz
• DOI: 10.1038/s44221-026-00583-3

Research Groups

• Institute of Hydrology and Water Management, BOKU University, Vienna, Austria
• baseflow AI solutions, Vienna, Austria

Short Summary

This study demonstrates the use of variational autoencoders (VAEs) as text-generating AI models to automatically derive interpretable parameter transfer functions for distributed hydrological and land-surface models. This novel approach significantly improves runoff predictions in ungauged basins across 162 German catchments compared to traditional regionalization methods and deep learning models, while maintaining physical interpretability.

Objective

• To automatically derive interpretable parameter transfer functions relating catchment attributes to hydrological model parameters using text-generating AI, thereby improving parameter estimation in data-scarce regions.

Study Configuration

• Spatial Scale: 162 mesoscale river basins across Germany.
• Temporal Scale: Daily meteorological forcings and river discharge data up to 2019.

Methodology and Data

• Models used:
  • Hydrological Model: mesoscale Hydrological Model (mHM version 5.11)
  • AI Model: Variational Autoencoders (VAEs) used as text-generating models for equation discovery.
• Data sources:
  • Meteorological forcings (daily precipitation, minimum, mean, maximum air temperature) and physio-geographical inputs (elevation, slope, aspect, soil properties, land cover) from a dataset published in ref. 56 (available at https://www.ufz.de/index.php?en=41160).
  • River discharge data from the Global Runoff Data Centre (https://grdc.bafg.de) and German state online portals.
  • Monthly leaf area index (LAI) from MODIS (ref. 58).
  • Gridded monthly climate data (precipitation, temperature, temperature range) from the Deutscher Wetterdienst (ref. 59).

Main Results

• The VAE-derived transfer functions led to improved runoff predictions in a prediction-in-ungauged-basins setting across 162 German basins.
• Performance was superior compared to established regionalization methods and regional long short-term memory (LSTM) networks.
• The learned functions demonstrated robustness across catchments and scalability to large spatial domains.
• The derived transfer functions maintained physical interpretability, addressing a key challenge in AI-based environmental modeling.

Contributions

• Introduction of a novel method using variational autoencoders (VAEs) as text-generating models to automatically discover interpretable parameter transfer functions for hydrological and land-surface models.
• Reformulation of equation discovery as an optimization problem in a continuous latent space, enhancing efficiency and transparency in parameter estimation.
• Demonstration of improved runoff prediction performance in ungauged basins compared to state-of-the-art regionalization and deep learning approaches.
• Providing a pathway towards more transparent, transferable, and physically interpretable parameter estimation for large-scale process-based environmental models.

Funding

• Funding was acquired by K.S. (Karsten Schulz). Specific project or program details are not provided in the paper text.

Citation

@article{Feigl2026Distilling,
  author = {Feigl, Moritz and Herrnegger, Mathew and Schulz, Karsten},
  title = {Distilling hydrological and land-surface model parameters from physio-geographical properties using text-generating AI},
  journal = {Nature Water},
  year = {2026},
  doi = {10.1038/s44221-026-00583-3},
  url = {https://doi.org/10.1038/s44221-026-00583-3}
}