ADOMBI (2026) DeepDiscover: towards autonomous discovery of bucket-type conceptual models – a proof of concept applied to hydrology

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-03-05
• Authors: Adoubi Vincent De Paul ADOMBI
• DOI: 10.1016/j.jhydrol.2026.135249

Research Groups

• Research Group R2Eau, Centre d’études sur les ressources minérales, Université du Québec à Chicoutimi, Canada

Short Summary

This study introduces DeepDiscover, a physics-embedded machine learning framework designed to autonomously infer bucket-type conceptual hydrological models from data. It demonstrates the feasibility and superior predictive performance of this approach compared to traditional benchmarks, reducing reliance on expert-defined model formulations.

Objective

• Can DeepDiscover achieve competitive predictive performance compared to existing conceptual, physics-embedded machine learning (PeML), and deep learning models?
• Can the framework generate hydrologically meaningful process and state variables without being provided with their explicit formulations?
• Does DeepDiscover exhibit physically coherent behavior when subjected to controlled perturbations in the input variables?

Study Configuration

• Spatial Scale: 569 river basins across the contiguous United States, selected from the CAMELS-US dataset.
• Temporal Scale: Daily meteorological variables and streamflow data. Warm-up period: 365 days. Training period: 1980-10-01 to 1999-09-30 (with 1993-10-01 to 1999-09-29 for validation). Testing period: 1999-10-01 to 2010-09-30.

Methodology and Data

• Models used: DeepDiscover-based PeML (DD-PeML), EXP-HYDRO (conceptual model), EXP-PeML (PeML incorporating EXP-HYDRO), Long Short-Term Memory (LSTM), and 1D Convolutional Neural Network (1D-CNN).
• Data sources: CAMELS-US dataset. Input data included daily precipitation, average temperature, and day length. Streamflow was the sole output target. Catchment attributes (27 variables) were also used for regional modeling. Potential evapotranspiration was estimated using Hamon’s equation.

Main Results

• Predictive Performance: DD-PeML (DD3-PeML configuration) achieved the highest overall predictive performance, with median Nash-Sutcliffe Efficiency (NSE) of 0.68 and median Kling-Gupta Efficiency (KGE) of 0.70 on the test set. It outperformed EXP-PeML (median NSE 0.62), 1D-CNN and EXP-HYDRO (median NSE around 0.54), and LSTM (median NSE 0.45). Performance was lower in arid central US regions for all models.
• Process Simulation and Discovery: DD-PeML (DD1-PeML configuration) learned internal process and state variables that were functionally consistent with the EXP-HYDRO model, showing median Pearson correlation coefficients of approximately 70 % for inferred processes and 80 % for state variables. An exploratory analysis with three candidate processes demonstrated the framework's ability to generate distinct, hydrologically meaningful components (e.g., snowmelt-driven, baseflow-like processes).
• Physical Coherence: Perturbation experiments confirmed that DD-PeML exhibited physically coherent responses to changes in precipitation and temperature. Increased precipitation led to increased streamflow and storage, while increased temperature (and thus evapotranspiration) led to reductions. Precipitation exerted a stronger influence (1.3 to 2.0 times greater) on state and streamflow than temperature.

Contributions

• Introduces DeepDiscover, a novel framework for the autonomous discovery of catchment-specific conceptual hydrological model structures, moving beyond predefined models.
• Advances physics-embedded machine learning by enabling the discovery of governing equations and process representations from data, rather than just parameter estimation.
• Demonstrates the feasibility of autonomously inferring bucket-type conceptual hydrological models within a physically constrained learning framework, reducing dependence on human expert-defined formulations.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Citation

@article{ADOMBI2026DeepDiscover,
  author = {ADOMBI, Adoubi Vincent De Paul},
  title = {DeepDiscover: towards autonomous discovery of bucket-type conceptual models – a proof of concept applied to hydrology},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135249},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135249}
}