The Role of Artificial Intelligence in Predicting Potentially Harmful Element Behaviour During Flooding Using Physicochemistry

Abstract

Anthropogenic contamination from the extensive use of potentially harmful elements (PHEs) in industrial, commercial, and agricultural activities is a significant public health concern. PHEs, such as lead, can become immobilised within soils, but are remobilised during flooding, which erodes soil and transports these elements downstream, depositing them on floodplains - often densely populated areas. Flooding alters porewater physicochemistry (e.g., pH), affecting PHE solid-phase distribution and bioaccessibility. Understanding these changes is crucial for managing the risks of increased exposure to remobilised PHEs during flooding. This thesis developed a machine-learning predictive tool to monitor changes in PHE porewater solubility, solid-phase distribution, and bioaccessibility during flooding. The tool was trained on physicochemical data from microcosm and mesocosm flood experiments and demonstrated the critical influence of soil particle size, redox potential, pH, and dissolved organic carbon on PHE dynamics. Decision tree models, particularly random forests, provided highly accurate predictions for PHE behaviour during flooding. When integrated with geographical information systems (GIS), these models enabled rapid, large-scale estimations of PHE changes, reducing the need for resource-intensive laboratory analyses. The findings suggest that machine-learning, informed by physicochemical data, offers a scalable and reliable method for predicting PHE dynamics during flooding. This approach may significantly support policymakers in identifying areas at risk of contamination under future flood scenarios. Future research should validate the random forest predictions across diverse catchments with varying physicochemical conditions.