Kayhomayoon et al. (2026) Developing the fusion of MODFLOW simulation and data-driven approaches for river-aquifer recharge modeling

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-02-19
• Authors: Zahra Kayhomayoon, Nazanin Sadeghi, Sami Ghordoyee Milan, Hannu Marttila, Naser Arya Azar
• DOI: 10.1016/j.jhydrol.2026.135169

Research Groups

• Department of Geology, Payame Noor University, Tehran, Iran
• Water Engineering Department, Faculty of Agricultural Technology, University College of Agriculture & Natural Resources, University of Tehran, Tehran, Iran
• Water, Energy and Environmental Engineering Research Unit, University of Oulu, Finland
• Department of Water Engineering, Faculty of Agriculture, University of Tabriz, Tabriz, Iran

Short Summary

This study developed a hybrid MODFLOW-machine learning approach to simulate and predict river-aquifer recharge in the Guilan aquifer, Iran, demonstrating its effectiveness for complex groundwater management and potential to reduce water loss by up to 30%.

Objective

• To develop and evaluate a hybrid approach combining MODFLOW numerical simulation with machine learning models (LSSVR, MARS, RF, GPR) for simulating and predicting river-aquifer recharge in the Guilan aquifer, Iran, aiming to reduce computational demands for long-term groundwater management.

Study Configuration

• Spatial Scale: Guilan aquifer, northern Iran, covering an area of approximately 1000 square kilometers (km²). The numerical model used a grid size of 500 meters (m) × 500 m, resulting in 1022 active cells.
• Temporal Scale: MODFLOW simulation included a steady state (October 2011) and a transient state (2011–2013). Groundwater level data for stability analysis spanned 28 years (1992–2019), with monthly fluctuations observed. Machine learning models were trained and tested using data derived from the 2011–2013 MODFLOW simulation outputs.

Methodology and Data

• Models used:
  • Numerical: MODFLOW 2000 (within the Groundwater Modeling System (GMS) 10.8 software environment).
  • Machine Learning: Multivariate Adaptive Regression Splines (MARS), Gaussian Process Regression (GPR), Least Squares Support Vector Regression (LSSVR), and Random Forest (RF).
  • Calibration: PEST method was employed for limited calibration of hydraulic conductivity and specific yield.
• Data sources:
  • Meteorological data (precipitation, evaporation).
  • Aquifer structural information (bedrock, topography).
  • Water use data.
  • Groundwater balance components (river information, exploitation wells, surface recharge).
  • Observations from 54 wells, with monthly groundwater level measurements.
  • Geological cross-sections, geophysical surveys, and well log data.
  • Regional Water Company of Guilan Province and Ministry of Energy reports.
  • Outputs from MODFLOW simulation, including simulated groundwater levels (GWL), top elevation, river heads, river beds, calibrated hydraulic conductivities, surface recharge rates, and river-aquifer exchange/recharge rates.

Main Results

• The MODFLOW numerical model successfully simulated aquifer-river interactions, with a Root Mean Squared Error (RMSE) of 0.84 m and Mean Absolute Error (MAE) of 0.77 m for the steady state, and RMSE of 1.19 m and MAE of 0.96 m for the transient state.
• Annual evaporation from the aquifer was simulated at approximately 15 × 10⁶ ± 1.5 × 10⁶ cubic meters (m³), and drainage at about 81 × 10⁶ ± 5 × 10⁶ m³.
• Implementing proper management strategies, such as controlled groundwater withdrawal, can reduce evaporation by 5 × 10⁶ to 8 × 10⁶ m³ per year and drainage by 15 × 10⁶ to 20 × 10⁶ m³ per year, potentially reducing total water loss by 20% to 25% (up to 30% as stated in the abstract).
• Among the machine learning models, Gaussian Process Regression (GPR) demonstrated the best performance in predicting river-aquifer recharge. For the test data, GPR achieved a Mean Squared Error (MSE) of 100 m³/month, RMSE of 9600 m³/month, MAE of 3700 m³/month, Nash–Sutcliffe Efficiency (NSE) of 0.9762, Scatter Index (SI) of 0.0917, and Willmott Index (WI) of 0.9938.
• Least Squares Support Vector Regression (LSSVR) also performed well, with results close to GPR. Multivariate Adaptive Regression Splines (MARS) provided a regression equation but showed lower performance than GPR and LSSVR. Random Forest (RF) exhibited the weakest predictive performance.
• The identified reliable input variables for the machine learning models were top elevation, groundwater level, and surface recharge to the aquifer.
• The hybrid MODFLOW-ML approach offers a significant reduction in computational time and resources for effective aquifer management.

Contributions

• Development of a novel hybrid framework integrating physically-based MODFLOW numerical modeling with data-driven machine learning (MARS, GPR, LSSVR, RF) for simulating and predicting complex river-aquifer recharge.
• Demonstration of the capability of ML models, particularly GPR, to accurately predict river-aquifer recharge based on MODFLOW outputs, thereby reducing the need for long-term, computationally intensive numerical simulations.
• Analysis of the performance of MARS and GPR models in conjunction with MODFLOW for groundwater resource management, highlighting their benefits in a hybrid approach.
• Identification of key input variables (topography, groundwater level, surface recharge) for efficient ML-based prediction of river-aquifer recharge.
• Quantification of significant water losses (evaporation, drainage) in the Guilan aquifer and proposal of management strategies (controlled withdrawal) to reduce these losses by up to 30%.
• The approach provides a transferable tool for evaluating river-aquifer recharge and supporting sustainable decision-making in complex and data-limited aquifer systems.

Funding

The paper mentions funding acquisition by Nazanin Sadeghiloyeh and Sami Ghordoyee Milan, but no specific projects, programs, or reference codes are provided.

Citation

@article{Kayhomayoon2026Developing,
  author = {Kayhomayoon, Zahra and Sadeghi, Nazanin and Milan, Sami Ghordoyee and Marttila, Hannu and Azar, Naser Arya},
  title = {Developing the fusion of MODFLOW simulation and data-driven approaches for river-aquifer recharge modeling},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135169},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135169}
}