What the study found
The study found that DeepDiscover can autonomously discover bucket-type conceptual hydrological models from data. The authors also report that the framework’s learned processes and states are consistent with EXP-HYDRO, a conceptual hydrological model.
Why the authors say this matters
The authors conclude that this work is a step toward reducing dependence on expert-defined model formulations. The study suggests that a physics-embedded machine learning framework can support data-driven process discovery in hydrology.
What the researchers tested
The researchers developed DeepDiscover, a modular neural architecture whose elementary units are intended to represent reservoirs in bucket-type conceptual hydrological models. They evaluated it on the CAMELS-US dataset in streamflow prediction using three experiments: benchmark comparison, recovery of EXP-HYDRO-like internal dynamics, and perturbation tests.
What worked and what didn't
The DeepDiscover-based model, called DD-PeML, outperformed EXP-HYDRO, EXP-PeML, a 1D-CNN, and an LSTM on the test set, with median NSE of 0.68 and median KGE of 0.70. When trained to mirror EXP-HYDRO, the inferred processes and states closely matched EXP-HYDRO, with median R2 of about 70% for processes and 80% for states, and perturbation experiments showed physically coherent responses to precipitation and temperature changes.
What to keep in mind
The abstract describes this as a proof of concept, so the findings are presented within that scope. Limitations beyond the use of CAMELS-US and the stated experiments are not described in the available summary.
Key points
- DeepDiscover autonomously infers bucket-type conceptual hydrological models from data.
- DD-PeML outperformed EXP-HYDRO, EXP-PeML, a 1D-CNN, and an LSTM on the CAMELS-US test set.
- The model achieved median test NSE of 0.68 and median KGE of 0.70.
- Inferred processes and states closely matched EXP-HYDRO when trained to mirror it.
- Perturbation experiments produced physically coherent responses to precipitation and temperature changes.
Disclosure
- Research title:
- DeepDiscover autonomously infers bucket-type hydrological models
- Authors:
- Adoubi Vincent De Paul Adombi
- Institutions:
- Université du Québec à Chicoutimi
- Publication date:
- 2026-03-05
- OpenAlex record:
- View
Get the weekly research newsletter
Stay current with scholarly research without reading academic papers — one filtered digest, every Friday.