Developing the fusion of MODFLOW simulation and data-driven approaches for river-aquifer recharge modeling

Abstract

A combination of groundwater numerical modeling with MODFLOW, and data-driven machine learning (ML) algorithms, including multivariate adaptive regression splines (MARS), Gaussian process regression (GPR), least squares support vector regression (LSSVR), and random forest (RF), were employed to simulate the interactions between rivers and aquifers within the Guilan aquifer, located in northern Iran. First, groundwater simulation was performed using MODFLOW. The river-aquifer recharge values were then calculated in various river routes. Afterwards, the ML models were developed based on the results of the MODFLOW simulation, where the river-aquifer recharge values calculated were considered targets. The model inputs included topography, groundwater level, and surface recharge to the aquifer, based on available datasets. The numerical model showed the capability to simulate the aquifer-river interactions, while among the ML methods, the GPR was capable of predicting the aquifer recharge and aquifer-river interactions with promising performance with MSE, RMSE, MAE, NSE, SI, and WI of 0.0001 MCM/month, 0.0096 MCM/month, 0.0037 MCM/month, 0.9762, 0.0917, and 0.9938, respectively, for the test data. Furthermore, the MODFLOW-ML were utilized to assess various management programs of the aquifer. The results showed that proper management programs can reduce the evaporation from the aquifer and drainage by up to 30%. The method presented in this study can be an effective approach in utilizing numerical models and ML to model complex aquifer-river recharge with reduced computational time and resources for more appropriate aquifer management.