From conceptual site models to data-driven insights: Key controls on vapor extraction performance in dual-phase extraction systems.

Conceptual site models (CSMs) are commonly used to support remediation decision-making in dual-phase extraction (DPE) systems, but their application often relies on qualitative judgment and site-specific experience. This study developed a data-driven framework to translate CSM-based interpretation into a quantitative form for remediation adjustment support. A total of 421 records from five DPE systems across three contaminated sites were compiled, and 13 input features were used to predict off-gas PID concentration. Three nonlinear machine-learning models, multilayer perceptron (MLP), radial basis function network (RBFN), and kernel ridge regression (KRR), were compared with a multiple linear regression (MLR) baseline. The nonlinear models consistently outperformed MLR, with KRR achieving the best test performance in both the full-feature (R2=0.73) and SHAP-based Top-10 reduced-feature (R2=0.75) settings, whereas MLR yielded lower values (R2=0.50 and 0.48). SHAP analysis was further used to interpret feature importance and guide reduced-feature model development, from which KRR_SET3 was identified as the most informative five-feature structure. Leave-one-site-out (LOSO) evaluation showed that reliable quantitative generalization across sites was not achieved, with negative external-site R2 values for all three left-out sites. However, when KRR_SET3 and KRR_all were applied to two unseen-site cases, KRR_SET3 more clearly captured upward concentration trends under changed conditions. These results suggest that, although precise cross-site prediction remains difficult, interpretable machine learning can preserve trend-sensitive signals useful for field interpretation and action-oriented support.