An explainable machine learning surrogate framework for colloid-facilitated contaminant transport in porous media.

Colloids act as important transport carriers in subsurface environments, facilitating the migration of strongly sorbing contaminants and significantly influencing breakthrough curve (BTC) in porous media. This study develops a machine-learning surrogate model capable of accurately reproducing BTCs generated by an equilibrium colloid-facilitated contaminant transport model over a wide range of transport conditions. The mechanistic model was first validated against the experimental phenanthrene transport data, demonstrating excellent agreement under both colloid-free and colloid-present conditions. The presence of colloids significantly accelerated contaminant transport, reducing the retardation factor from 339.69 to 64.76. Three machine-learning algorithms, namely Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM), were evaluated for BTC prediction. Among them, RF exhibited the highest predictive accuracy with an R2 of 0.9982, RMSE of 1.23×10-2, and MAE of 5.89×10-3. Explainable artificial intelligence analyses based on feature importance and SHapley Additive exPlanations (SHAP) revealed that flow velocity was the dominant controlling parameter, followed by contaminant partitioning coefficients associated with mobile and immobile colloids. Furthermore, the RF surrogate accurately reproduced BTCs across diverse transport scenarios while maintaining excellent generalization capability. The novelty of this study lies in the integration of a physics-based colloid-facilitated contaminant transport model with interpretable machine learning and explainable artificial intelligence to develop a computationally efficient surrogate framework that not only accurately predicts BTCs but also reveals the relative importance of governing transport parameters.