Abstract
A machine-learning-based methodology is proposed to delineate the spatial distribution of geomaterials across a large-scale three-dimensional subsurface system. The study area spans the entire Po River Basin in northern Italy. As uncertainty quantification is critical for subsurface characterization, the methodology is specifically designed to provide a quantitative evaluation of prediction uncertainty at each location of the reconstructed domain. The analysis is grounded on a unique dataset that encompasses lithostratigraphic data obtained from diverse sources of information. A hyperparameter selection technique based on a stratified cross-validation procedure is employed to improve model prediction performance. The quality of the results is assessed through validation against pointwise information and available hydrogeological cross-sections. The large-scale patterns identified are in line with the main features highlighted by typical hydrogeological surveys. Reconstruction of prediction uncertainty is consistent with the spatial distribution of available data and model accuracy estimates. It enables one to identify regions where availability of new information could assist in the constraining of uncertainty. The comprehensive dataset provided in this study, complemented by the model-based reconstruction of the subsurface system and the assessment of the associated uncertainty, is relevant from a water resources management and protection perspective. As such, it can be readily employed in the context of groundwater availability and quality studies aimed at identifying the main dynamics and patterns associated with the action of climate drivers in large-scale aquifer systems of the kind here analyzed, while fully embedding model and parametric uncertainties that are tied to the scale of investigation.