Artificial neural networks vis-à-vis MODFLOW in the simulation of groundwater: a review

Although numerical and non-numerical models of groundwater flow and transport have separately been reviewed in several studies, they have not hitherto been reviewed simultaneously. Additionally, few case studies have considered these two models to simulate groundwater. The purpose of this study is to compare MODFLOW and artificial neural networks (ANNs) as the most typical numerical and non-numerical groundwater models, respectively, with placing the emphasis on the review of studies in which both models have been considered. Until the previous decade, MODFLOW was quantitatively used far more than ANNs to simulate groundwater. However, since then, the application of ANNs in groundwater has significantly augmented in comparison with MODFLOW. A thorough understanding of the physical properties of the aquifer along with having accurate and sufficient data are requisite to simulate groundwater using MODFLOW. Moreover, despite existing automatic calibration methods, e.g. PEST, MODFLOW is ordinarily calibrated by trial and error, which is onerous and time-consuming. This model is typically applied to alluvial aquifers, which are assumed to be homogeneous and isotropic. On the other hand, ANNs with a black box approach can simulate groundwater through data excluding aquifer's characteristics, e.g. through utilizing the climatic variables. Therefore, ANNs may straightforwardly be applied to the heterogeneous and anisotropic aquifers, i.e. karst and hard-rock. However, determining the dynamic response of the aquifer may be of central importance despite the formidable challenges related to the application of numerical models. Therefore, they have been selected to simulate the response of the complicated aquifers, especially in recent studies.