Vapor Intrusion Site Characterization under Fluctuating Groundwater Tables: A Machine Learning-Enhanced Ensemble Smoother Framework

Accurate characterization of subsurface dynamics is critical for assessing soil vapor intrusion (VI) risks under fluctuating groundwater tables (GWT). While the ensemble smoother (ES) integrates model simulations with observational data, it can be challenged by high-dimensional nonlinear systems. To address this limitation, we propose a machine learning-enhanced ES framework (ESML) that replaces the conventional parametric ES update with nonparametric update operators learned using integrated machine learning, including random forest (RF), artificial neural networks (ANN), and support vector regression (SVR). Results demonstrate that the RF-based ESML (ESRF) achieves the best inversion accuracy (RMSE = 0.30 ± 0.02), outperforming ESANN and ESSVR by 14–33% while resolving spatially heterogeneous soil’s properties. Large-amplitude GWT fluctuations (2 m amplitude) improve inversion accuracy by ∼20%, whereas the fluctuation period has a minor effect (∼5% variation). Furthermore, under moderate observational errors (σ ≤ 0.01), doubling soil water content observations can substitute vapor concentration data with comparable precision. Collectively, ESML provides a cost-effective pathway for VI characterization in data-limited sites with dynamic hydrological regimes.