Comparison of a Semi-Physical Model and an Artificial Neural Network for Simulating Past Groundwater Levels, in a Karst Aquifer in Southern France
Abstract
Because of their strongly heterogeneous permeability, karst aquifers exhibit nonlinear relations between their outputs (discharge or groundwater level) and inputs (rain, temperature, etc.). Three main types of models are commonly used to simulate karst aquifers: physical models, pipe-flow models and conceptual models. The “Méthode d’Évaluation des Crues Karstiques” or Method for Assessing Karst Floods (MECK) was implemented to perform the simulation of Cadarache’s fractured karst aquifer’s level (south of France). It relies on a semi-physical model that conceptualizes the functioning of the aquifer. Even though it simulates a continuous daily time series, the model’s optimization focuses on high levels for enhanced flood representation. Although MECK shows good performances on high groundwater levels, it encounters difficulties under specific conditions. The model struggles to accurately reproduce level variations caused by recent rainfall events. Despite having a large number of parameters, this semi-physical model is not able to reproduce the complex phenomena that govern the response of a karst aquifer. Artificial Neural Networks, however, are able to learn the system’s behavior directly from the data, without the operator needing. Most neural networks developed in the field of hydro(geo)logy are used to simulate spring discharge based on rain. Fewer studies focus on simulating groundwater levels, and among those, only a small subset specifically addresses karst aquifers. The Cadarache site belonging to the French Atomic Energy and Alternative Energies Commission (CEA) in southeastern France is subject to rapid kinetic groundwater flooding (several tens of meters rise in a few hours). With the aim of estimating the 100-year flood level for enhancing the safety of nuclear facilities, the goal of this study is to compare the performance of MECK against a Neural Network in simulating past groundwater level of a fractured karst aquifer, with a particular emphasis on high levels related to flooding. Daily groundwater levels are simulated from 1960 to 2025, using meteorological data. The data are pre-processed based on cross correlation analysis and stationarity tests. The neural network, a Multilayer Perceptron, is then developed focusing on the high levels of the groundwater. Its architecture is selected through a trial-and-error approach, and the hyperparameters are selected by optimizing cross-validation metrics. Simulated groundwater levels of MECK and the developed Neural Network are finally compared using classical hydrological metrics (NSE, RMSE, etc.), both for the entire time series, and for high levels. Results show that the Neural Network performs overall better than the semi-physical model. Whereas the levels simulated by the semi-physical model are mostly higher than the measured ones, the Artificial Neural Network reproduces accurately the overall dynamic and shows good performances on high levels but struggles to reproduce intermediate levels.
Related
- marineGeochemFoam: A multipurpose framework for simulating reactive transport and biogeochemical cycling in marine sediments and at seawater-groundwater interfaces
- Groundwater Potential Mapping in Northwestern Nigeria: A Comparison of the Analytic Hierarchy Process and Galileo-Enhanced CAT-XGB
- A data-driven method coupling multiple physical constraints for correcting structural errors in groundwater contaminant transport models
- Multivariate phase-space interpretation of ensemble forecasts: A comparison of physical and AI models
- Artificial intelligence/machine learning estimation of FAO and ASCE standardized reference ET erodes the physical basis and intended purpose of ET standardization
BibTeX
@article{jouanneaucouvret2026comparison,
title = {{Comparison of a Semi-Physical Model and an Artificial Neural Network for Simulating Past Groundwater Levels, in a Karst Aquifer in Southern France}},
author = {Sayane Jouanneau-Couvret and Jose Manuel Tunqui Neira and Sébastien Morilhat and Séverin Pistre and Anne Johannet},
journal = {Inżynieria Mineralna},
year = {2026},
month = nov,
doi = {10.29227/im-2026-02-02-014},
url = {https://imt-mines-ales.hal.science/hal-05857390v1},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Comparison of a Semi-Physical Model and an Artificial Neural Network for Simulating Past Groundwater Levels, in a Karst Aquifer in Southern France
AU - Sayane Jouanneau-Couvret
AU - Jose Manuel Tunqui Neira
AU - Sébastien Morilhat
AU - Séverin Pistre
AU - Anne Johannet
PY - 2026
DA - 2026/11/01
JO - Inżynieria Mineralna
DO - 10.29227/im-2026-02-02-014
UR - https://imt-mines-ales.hal.science/hal-05857390v1
AB - Because of their strongly heterogeneous permeability, karst aquifers exhibit nonlinear relations between their outputs (discharge or groundwater level) and inputs (rain, temperature, etc.). Three main types of models are commonly used to simulate karst aquifers: physical models, pipe-flow models and conceptual models. The “Méthode d’Évaluation des Crues Karstiques” or Method for Assessing Karst Floods (MECK) was implemented to perform the simulation of Cadarache’s fractured karst aquifer’s level (south of France). It relies on a semi-physical model that conceptualizes the functioning of the aquifer. Even though it simulates a continuous daily time series, the model’s optimization focuses on high levels for enhanced flood representation. Although MECK shows good performances on high groundwater levels, it encounters difficulties under specific conditions. The model struggles to accurately reproduce level variations caused by recent rainfall events. Despite having a large number of parameters, this semi-physical model is not able to reproduce the complex phenomena that govern the response of a karst aquifer. Artificial Neural Networks, however, are able to learn the system’s behavior directly from the data, without the operator needing. Most neural networks developed in the field of hydro(geo)logy are used to simulate spring discharge based on rain. Fewer studies focus on simulating groundwater levels, and among those, only a small subset specifically addresses karst aquifers. The Cadarache site belonging to the French Atomic Energy and Alternative Energies Commission (CEA) in southeastern France is subject to rapid kinetic groundwater flooding (several tens of meters rise in a few hours). With the aim of estimating the 100-year flood level for enhancing the safety of nuclear facilities, the goal of this study is to compare the performance of MECK against a Neural Network in simulating past groundwater level of a fractured karst aquifer, with a particular emphasis on high levels related to flooding. Daily groundwater levels are simulated from 1960 to 2025, using meteorological data. The data are pre-processed based on cross correlation analysis and stationarity tests. The neural network, a Multilayer Perceptron, is then developed focusing on the high levels of the groundwater. Its architecture is selected through a trial-and-error approach, and the hyperparameters are selected by optimizing cross-validation metrics. Simulated groundwater levels of MECK and the developed Neural Network are finally compared using classical hydrological metrics (NSE, RMSE, etc.), both for the entire time series, and for high levels. Results show that the Neural Network performs overall better than the semi-physical model. Whereas the levels simulated by the semi-physical model are mostly higher than the measured ones, the Artificial Neural Network reproduces accurately the overall dynamic and shows good performances on high levels but struggles to reproduce intermediate levels.
SN - 1640-4920
ER -