Conceptual Model of Arsenic Mobility in the Shallow Alluvial Aquifers Near Venice (Italy) Elucidated Through Machine Learning and Geochemical Modeling

This work proposed a novel method to elucidate the controls of As mobility in complex aquifersbased on an unsupervised machine learning algorithm, Self-Organizing Map (SOM), and process-basedgeochemical modeling. The approach is tested in the shallow aquifers of the Venetian Alluvial Plain (VAP)near Venice, Italy, where As concentrations seasonally and locally exceed recommended drinking waterlimits. SOM was fed using information from two geochemical surveys on eight VAP boreholes, andcontinuous reading of hourly groundwater head levels and weekly geochemical analyses from three VAPboreholes between mid-October 2017 and end of January 2018. The SOM analysis is consistent withredox-controlled dissolution-precipitation hydrous ferric oxides (HFOs) as a key control of As mobility in theaquifer. Dissolved As is positively correlated to Fe and NHþ4 and negatively to the oxidizing-reducingpotential (ORP). Negative correlation between As and groundwater head levels suggests a redox control byrainfall-driven recharge, which adds oxidants to the aquifer while progressively attenuating As. Thismechanism is tested using process-based geochemical modeling, which simulates different transportmodalities of oxidants entering the aquifer. Starting from reducing aquifer conditions, the model reproducescorrectly the observed ORP and the trends in As and Fe, when the function describing the occurrence ofoxidizing events scales according to the temporal occurrence of rainfall events. Heterogeneity can stronglycontrol the local-scale effectiveness of recharge as a natural As attenuating factor, requiring a differentmodel analysis to be properly assessed and to be developed in a follow-up study.