Comparison of physics-based and machine learning groundwater models for sustainable aquifer management: A regional study in Emilia-Romagna (Italy)

Study region 7000 km 2 portion of Emilia-Romagna region, north-eastern Italy. Groundwater resources are stored in a complex multi-layered aquifer system. Study focus A MODFLOW 6 numerical groundwater flow model and a random forest algorithm are implemented to compare the performance of a physics-based and a machine learning approach in simulating and predicting historical and future groundwater head values in the same study area. A scenario analysis is carried out to assess the response of the regional multi-layered aquifer system to variations in climatic and anthropogenic stresses. New hydrological insights for the region The MODFLOW 6 model and the random forest algorithm produce consistent results, although the random forest is generally more optimistic and predicts a lower impact of precipitation and pumping changes on the regional aquifer system. Both models indicate that future reduced precipitation combined with increased groundwater withdrawal would negatively affect shallow and deep aquifers. Conversely, decreasing pumping can partially counterbalance lower precipitation, allowing groundwater heads to recover in some areas. The analysis also identifies areas most sensitive to perturbations, where precipitation and abstraction variations lead to the largest groundwater head declines. Results emphasize the advantages of integrating physics-based and machine learning methods to better interpret outcomes and improve model performance. This approach strengthens both the calibration process and scenario analyses, offering a valuable contribution given the growing relevance of groundwater modeling for water management and policy decisions.