Decoding Arsenic Mobility in Soils: Machine Learning Insights into pH-Organic Carbon Interactions Across Global Texture Classes

Abstract Understanding the factors that govern arsenic mobility in soils is essential for effective risk assessment and remediation. Despite existing knowledge, a global-scale synthesis linking soil properties to arsenic mobility across diverse texture classes remains limited. This study addresses this gap by using a globally harmonized dataset and machine learning (ML) models, the combined influence of soil pH, organic carbon (OC), cation exchange capacity, total arsenic and clay content on the mobile arsenic fraction (PF₁) has been modelled. Random Forest, XGBoost, and Generalized Additive Models were trained to predict PF₁, with Random Forest outperforming others (R² = 0.65, RMSE = 13.21 and MAE = 6.93). Simulation involving 25,200 data points across 12 USDA texture classes revealed pronounced non-linear interactions between pH and OC, significantly affecting PF₁ values. Arsenic mobility was found to be lowest under slightly acidic pH (5.5-6.5) and low-to-moderate OC levels (< 2%), particularly in clay-rich soils. Conversely, coarse-textured soils exhibited broader “mobility control zones” where PF₁ remained below the 10% risk threshold, while fine-textured soils were more sensitive to pH and OC fluctuations. Kernel density estimations further delineated typical OC-pH combinations associated with low arsenic mobility. Although redox conditions and mineralogical influences were beyond the scope of this study, the findings underscore the importance of incorporating physicochemical parameters and soil texture into arsenic risk frameworks. Attempts of reducing arsenic concentrations through amendments or stabilization, management practices should also account for soil texture and chemistry, ensuring interventions are tailored to minimize the fraction of arsenic that remains mobile. Graphical Abstract This graphical abstract provides a concise, visually engaging summary of the research, enabling readers to rapidly grasp the study’s key insights without delving into the full manuscript. The central focus of the study is the prediction of arsenic (As) mobility in soils using global harmonized datasets incorporating critical soil attributes such as organic carbon (OC), soil pH, cation exchange capacity, clay content, total arsenic in soil, and mobile arsenic fraction (PF₁). These inputs feed into machine learning and statistical modeling approaches-Generalized Additive Models , XGBoost, and Random Forest-to assess and predict arsenic behaviour across diverse soil systems. Model performance indicators (R², RMSE, and MAE) are displayed for comparative evaluation, highlighting model efficiency. At the core of the abstract, a globe symbolizes the global relevance of the harmonized data, while simulation outputs illustrate 25,200 data points across 12 soil texture classes to evaluate OC and pH effects. The non-linear 3D surface plot emphasizes complex relationships between arsenic fraction, OC, and pH, reinforcing the need for advanced modelling. Furthermore, the lower left quadrant illustrates that fine-textured soils exhibit greater sensitivity to pH and OC fluctuations, while coarse-textured soils with high OC content tend to enhance arsenic mobility. This visualization underscores the importance of considering soil type and chemical properties in environmental risk assessments. Collectively, the graphical abstract encapsulates the study’s methodology, findings, and implications for managing arsenic contamination in agricultural soils at a global scale.