The Modeling of a River Impacted with Tailings Mudflows Based on the Differentiation of Spatiotemporal Domains and Assessment of Water-Sediment Interactions Using Machine Learning Approaches

The modeling of metal concentrations in large rivers is complex because the contributing factors are numerous, namely, the variation in metal sources across spatiotemporal domains. By considering both domains, this study modeled metal concentrations derived from the interaction of river water and sediments of contrasting grain size and chemical composition, in regions of contrasting seasonal precipitation. Statistical methods assessed the processes of metal partitioning and transport, while artificial intelligence methods structured the dataset to predict the evolution of metal concentrations as a function of environmental changes. The methodology was applied to the Paraopeba River (Brazil), divided into sectors of coarse aluminum-rich natural sediments and sectors enriched in fine iron- and manganese-rich mine tailings, after the collapse of the B1 dam in Brumadinho, with 85-90% rainfall occurring from October to March. The prediction capacity of the random forest regressor was large for aluminum, iron and manganese concentrations, with average precision > 90% and accuracy < 0.2.