Assessing groundwater sustainability through physics-based and machine learning models: a comparative study in Emilia-Romagna region (Italy)

Groundwater plays a crucial role in sustaining human population, agriculture, and ecosystems – particularly in arid and drought-prone areas. It provides a dependable freshwater source in periods of surface water shortage, especially during long-term drought events. However, as climate change has driven an increase in the frequency and severity of drought events, there is growing concern about the resilience of groundwater systems. These systems can take a long time to recover after a drought event. Therefore, studying the robustness and dynamics of these aquifer systems is crucial to safeguard fresh groundwater resources in regions prone to drought events. Due to this, sustainable groundwater management is particularly critical under changing climate conditions, as balancing freshwater demand with availability grows more challenging. Groundwater modeling is essential for gaining a deeper understanding of the current state of aquifer systems and for estimating future groundwater availability and trends under various scenarios. This knowledge supports more effective and sustainable groundwater and integrated freshwater management. In this study, we implement both a numerical groundwater flow model and a random forest algorithm to compare the performance of a physics-based approach and a machine learning method in simulating historical and future groundwater head values in a portion of the Emilia-Romagna region. For both methods, calibration is performed using an extensive dataset of piezometric level observations over a 17-year period. The study area includes approximately 130 wells from the regional monitoring network, each providing two measurements per year. The numerical groundwater flow model is developed using MODFLOW 6. It builds on a previous MODFLOW application to the whole Emilia-Romagna region by the Regional Agency for Prevention, Environment and Energy of Emilia-Romagna (Arpae), and extends on a large area east of the Secchia River. The random forest algorithm is implemented using the same input data and groundwater head observations in the same study area. It mainly incorporates hydrogeological, climatic, and topographic variables, along with information about groundwater abstraction within the region. These two modelling approaches are then applied to conduct a scenario analysis under varying climatic and groundwater pumping conditions. This analysis serves three main purposes: (i) to assess the impact of reduced distributed groundwater recharge, (ii) to explore the combined effects of natural and artificial stresses on the regional aquifer system, and (iii) to compare groundwater head predictions from both modeling approaches under identical scenario conditions. Our results show that, in this specific case, the random forest algorithm more accurately reproduces historical groundwater head values; however, its lack of physical interpretability may limit the understanding of the underlying aquifer dynamics. In the scenario analysis, the random forest algorithm predicts less variable groundwater head values compared to the MODFLOW model (as shown in Figure 1), likely due to its limited physical grounding, which may reduce its capacity and reliability in generalizing the outcome to new conditions – an important consideration for climate adaptation planning. Despite this, the algorithm offers valuable insights into the relative importance of input variables, which can help interpret and enhance the numerical model performance. Overall, the results highlight the benefits of combining physics-based and machine learning methods to strengthen groundwater modeling and support the definition of effective strategies for sustainable aquifer management to improve the resilience of regional groundwater systems.