A comparative analysis of physics-based and machine learning methods for sustainable aquifer management in the Emilia-Romagna region (Italy)

The Emilia-Romagna region is located in north-eastern Italy and hosts extensive agricultural and industrial activity along with densely populated urban centers. All these elements contribute to increase the water demand, which often relies on groundwater resources, especially during droughts.

The complex regional aquifer system, consisting of multiple interconnected layers, presents a challenging yet compelling case study. Moreover, the region benefits from hydrogeological and environmental data gathered through long-term monitoring and research activities, offering a robust foundation for further detailed analysis.

In this study we estimate the potential evolution of groundwater conditions in part of the Emilia-Romagna region, considering the impacts of climate change and human activities. In particular, the goal is to evaluate the resilience of the regional multi-layered aquifer system to prolonged drought conditions, and to outline potential guidelines for long-term sustainable regional groundwater management. Two modeling techniques are employed: a numerical groundwater flow model and a random forest algorithm. This dual approach allows to compare the performance of a physics-based and a machine learning model in simulating historical and future groundwater levels within the same study area, thus investigating the potential benefits of combining both methods.

In the first phase, a groundwater model is implemented using MODFLOW 6, alongside a random forest algorithm developed in R. Input data are sourced from a MODFLOW model covering the entire Emilia-Romagna groundwater system by Arpae (Regional Agency for Prevention, Environment and Energy of Emilia-Romagna), as well as from publicly accessible datasets available through the Emilia-Romagna Region and Arpae repositories.

Next, we use the groundwater model and the random forest algorithm to analyze scenarios under different climatic and groundwater abstraction conditions. The aim is to assess the combined impacts of hypothetical drought events and changes in groundwater pumping rate regime on the groundwater heads in the regional aquifer system.  Results from both approaches suggest that the aquifer system is vulnerable to potential future droughts. While increased groundwater abstraction could intensify the effects of reduced precipitation, decreasing groundwater pumping might partially alleviate the drought effects. Specific areas are also pinpointed where the impacts of reduced precipitation, changes in pumping rate, or their combination are more significant. This underscores the importance of evaluating both the overall study region and local scales to identify critical hotspots and determine the most effective strategies for mitigation and adaptation to future droughts and climate change.

The random forest algorithm offers valuable insights into the relative importance of data and variables influencing the final groundwater head distribution, enhancing the interpretation of the groundwater model results and suggesting areas for potential improvement. However, due to its lack of physical interpretability, it presents a lower generalization capability compared to a numerical model. These findings highlight the advantages of integrating physics-based and machine learning approaches to understand model outputs and improve overall performance. Combining the two methods strengthens both the calibration process and the scenario analysis, providing a significant contribution to groundwater modeling, which will play an increasingly important role in the future.