Liang, Xiao Xia 
ORCID: https://orcid.org/0009-0007-5341-6762; Gloaguen, Erwan ORCID: https://orcid.org/0000-0002-9400-0276; Claprood, Maxime ORCID: https://orcid.org/0009-0002-2697-1166; Paradis, Daniel ORCID: https://orcid.org/0000-0003-0727-4541 et Lauzon, Dany ORCID: https://orcid.org/0000-0001-7774-4460
(2025).
Graph Neural Network Framework for Spatiotemporal Groundwater Level Forecasting.
Mathematical Geosciences
, vol. 57
, nº 6.
pp. 1071-1093.
DOI: 10.1007/s11004-025-10194-5.
Résumé
Spatiotemporal groundwater (GW) forecasting plays an important role in water resources management and planning. Given the increasing anthropogenic and climatic pressures on GW, there is an urgent need for more efficient and accessible forecasting tools to ensure the sustainable management and distribution of this vital resource. Traditional methods, including simple empirical relationships and complex numerical simulations, have shown varying success. However, these approaches often struggle to incorporate new information and are computationally demanding, which can compromise the accuracy and reliability of forecasts over time. To address these limitations, we propose a graph neural network (GNN)-based framework that integrates explicit geostatistical methods with implicit machine learning techniques, serving as a surrogate for numerical models to provide fast and accurate GW level forecasts. This GNN framework can be used as a coupling tool with the numerical model to aid in forecasting; it is not limited to surrogacy. GNNs excel in handling complex data structures, making them well suited for geosciences applications. Our framework is trained on simulated transient GW levels and pumping rates generated from a complex hydrogeological model in Quebec, Canada. It has been tested under various pumping scenarios, each taking several days to solve using a dedicated flow simulator. In contrast, our approach delivers results in seconds, offering valuable support for GW management under complex stress conditions. The results highlight the framework’s ability to provide rapid and reliable GW forecasts, offering a computationally efficient alternative for managing GW in intricate hydrogeological settings.
| Type de document: | Article |
|---|---|
| Mots-clés libres: | machine learning; deep neural network; graph convolutional network; long short-term memory; groundwater forecasting; surrogate models |
| Centre: | Centre Eau Terre Environnement |
| Date de dépôt: | 18 juin 2026 12:57 |
| Dernière modification: | 18 juin 2026 12:57 |
| URI: | https://espace.inrs.ca/id/eprint/16666 |
Gestion Actions (Identification requise)
 |
Modifier la notice |