Machine learning models for predicting surfactant-enhanced oil removal from contaminated soil

Abstract

Soil contamination is a significant environmental concern due to the persistence and toxicity of petroleum hydrocarbons. Surfactant-enhanced remediation (SER) enables effective removal of such pollutants by facilitating their solubilization and desorption. However, conventional linear and mechanistic models struggle to capture the complex nonlinear interactions among soil characteristics, pollutant types, and surfactant formulations. Machine learning provides a powerful alternative to address these challenges. In this work, a dataset of 2394 samples encompassing surfactant, soil, oil, water properties, and operational conditions was used to train and validate six predictive models. Among them, the categorical boosting (CB) model exhibited the highest performance (R² = 0.985 and RMSE = 0.068), followed by extreme gradient boosting and decision tree. Cross-validation confirmed the robustness of the CB model, with 96.4 % of predictions within the statistical applicability domain, while benchmarking demonstrated its ability to reproduce experimental trends. Monte Carlo based sensitivity analysis showed agitation speed, surfactant concentration, liquid-to-soil ratio, and washing time as the dominant factors controlling remediation efficiency. These findings demonstrate that AI-driven modeling can serve as an efficient tool for environmental remediation, enabling optimized surfactant-based soil cleanup designs while reducing operational risks and costs compared to conventional methods.