A data-mechanism integrated framework for elucidating the migration and attenuation of light non-aqueous phase liquids in the vadose zone.

Elucidating key driving factors and the underlying mechanisms governing the complex subsurface fate of Light non-aqueous phase liquids (LNAPL) is essential for effective pollution characterization and remediation. Here, we develop the "data-mechanism" integrated framework DFEMR (dataset development - feature evaluation - mechanistic revelation). The DFEMR framework synergizes sandbox simulations with data-driven analysis to identify key driving factors through feature importance and XGBoost-SHAP (eXtreme Gradient Boosting - SHapley Additive exPlanations) contributions, which are mechanistically validated by targeted controlled experiments. Results showed a "tripartite spatiotemporal pattern" of LNAPL in the vadose zone, with total petroleum hydrocarbon (TPH) and environmental variables exhibiting an "accumulation-breakthrough" trend. "Position-Y" and "Volumetric Water Content (VWC)" achieved leading positions in the driving factors' ranking, while pH and "oxidation-reduction potential (ORP)" was identified as the regulatory factors. Critically, controlled experiments provided mechanistic explanations for data-driven insights: VWC modulated LNAPL's sorption and partitioning, altered its dynamic equilibrium at varying depths while synergizing with pH and ORP to achieve optimal biodegradation amount (5707-9638 mg/kg) and rate (22.21-24.37%) at 10-15% levels. Pilot-scale application confirmed the framework's generalization capability. DFEMR enhances interpretability of data-driven models and overcomes variable prioritization constraints of conventional experiments, offering a holistic approach for elucidating LNAPL migration and attenuation.