Machine Learning Predicts Flood-Driven Porewater Concentrations and Bioaccessibility of Potentially Harmful Elements Across Contrasting Catchments

Abstract Climate change can intensify flooding, creating a pressing need for predictive tools to support environmental assessment and monitoring of changes in soil porewater concentrations and oral bioaccessibility of potentially harmful elements. Evidence for model transfer across geochemically contrasting catchments remains limited. We tested whether machine learning models trained in one catchment could predict these dynamics in another without recalibration. Acidic, organic-rich Clyde soils and alkaline, organic-poor Humber soils were analysed before and after flooding. Random forest, XGBoost, and neural networks were trained using microcosm data from flood-prone soils along the River Clyde. Random forest and XGBoost performed best in predicting porewater arsenic, lead, and copper concentrations (R² > 0.70) and gastric bioaccessibility (R² ≥ 0.89) in the Clyde. Four phases were identified: first lag, critical, subcritical, and second lag. Incorporating critical-phase data (days 2–4), when hydrochemical change peaked, improved porewater concentration predictions in the Clyde (ΔR² ≤ 0.13; RMSE reduction ≤ 15 µg/L). Clyde-trained models were applied to Humber soils; no Humber data were used for training, tuning, or recalibration. Random forest showed the strongest cross-catchment transfer to the Humber, with no significant differences between predicted and measured porewater copper or lead concentrations at any sampling point. Arsenic predictions differed at one sampling point (p = 0.022). Random forest predictions closely matched measured gastric and gastric-intestinal bioaccessibility. Targeting rapid hydrochemical change could help optimise sampling design for model training. Alongside this, cross-catchment model reuse could extend flood-related contamination and exposure assessment, supporting broader-scale environmental management of flood-prone contaminated soils.