Modelling groundwater dynamics for sustainable aquifer management: a comparative study of physics-based and machine learning approaches in the Emilia-Romagna region (Italy)

The sustainable use of groundwater poses a recognized global challenge for sustainable development. In the coming years, aquifer depletion and groundwater overexploitation may make it harder and harder for communities to address the growing demand and the reduced availability of water resources. The Emilia-Romagna region in Italy is an excellent case of a highly monitored aquifer system playing an essential role for water supply for civil, agricultural, and industrial use. An extended agricultural plain is located in this area, and large amounts of detailed information on aquifer characteristics, water withdrawals, and water table levels is available. The subsurface consists of multiple aquifers at different depths in fluvial sediment deposits of several hundred meters thickness in total, underlaid by marine sediment deposits. In this study, we implement both a numerical groundwater flow model and a random forest algorithm to compare the performance of a physics-based and a machine learning method in reproducing historical groundwater head values in a portion of the Emilia-Romagna region. In both cases, calibration is carried out by means of a wide dataset of piezometric levels observations over a 17-years time span. In the study area, about 130 observation wells are present that are part of the regional monitoring network, each providing two measures per year. The numerical groundwater flow model is developed in MODFLOW 6. It is based on a previous application of MODFLOW to the whole Emilia-Romagna area by the Regional Agency for Environmental Protection (ARPAE), and extends over a wide area east of the Secchia River. After the calibration phase, the final model setup implemented in this study provides good model performance, with a value of R2 coefficient equal to 0.89. In the calibration plot in Figure 1, the plane has been divided into regular hexagons, and each hexagon fill has been assigned according to the number of points it contains. The straight red line represents the equality of observed and simulated data, which outlines perfect model performances. The random forest algorithm is implemented banking on input data and groundwater head observations in the same study area. It mainly considers hydrogeological, climatic, and topographic variables, as well as information about groundwater extraction in the study area. The results of this study show that in this specific case the random forest algorithm provides more accurate performance in reproducing groundwater head observations. Even though the machine learning approach may lack physical interpretability, it can contribute to the identification of the most significant factors affecting groundwater systems, and therefore it can provide valuable insights of the effects of natural and artificial stress on groundwater levels in the Emilia-Romagna region. These, in turn, would provide guidelines for sustainable aquifer management to improve the resilience of regional aquifers.