Physics-Informed Predictive Analytics for Subsurface Fluid–Contaminant Migration Using Hybrid Geophysical Inversion and Machine Learning

Abstract

Geological heterogeneity poses a major challenge to predicting subsurface fluid–contaminant migration and assessing groundwater vulnerability. This study develops a physics-informed predictive framework that integrates geophysical inversion with machine learning (ML) to improve the prediction and interpretation of contaminant migration in heterogeneous subsurface environments. A synthetic dataset comprising 10,000 subsurface realizations was generated within physically representative ranges of electrical resistivity (10–1200 Ωm), seismic velocity (2.5–5.8 km/s), induced polarization chargeability (5–35 mV/V), porosity (12–32%), permeability (80–520 mD), and fracture connectivity (0.2–0.9). The dataset was partitioned into 7,000 training, 1,500 validation, and 1,500 independent test samples, with a fixed random seed used to ensure reproducibility. Geophysical and hydrogeological variables were integrated through constrained modelling and used as inputs to Random Forest (RF) and Gradient Boosting (GB) classifiers for contaminant-migration risk prediction. Model performance was evaluated using accuracy, precision, recall, and F1-score, while feature-importance analysis was used to identify the dominant controls on predicted contaminant migration. The class-specific evaluation produced accuracies ranging from 76.9% to 84.6% and F1-scores from 0.75 to 0.84 across the three contaminant-risk classes. In the ablation analysis, the complete physics-informed configuration achieved independent-test accuracies of 91.07% for Random Forest and 91.47% for Gradient Boosting, with corresponding macro-F1 scores of 0.900 and 0.904, respectively. Relative to the conventional machine-learning baseline, the Gradient Boosting macro-F1 increased from 0.771 to 0.904, representing a 17.3% relative improvement under identical test conditions. Resistivity, permeability, and fracture connectivity emerged as important predictors of contaminant-migration behaviour, reflecting the influence of fluid saturation, hydraulic properties, and preferential flow pathways. The results demonstrate the potential of integrating geophysical information with machine learning to provide physically interpretable predictions of contaminant-migration risk in heterogeneous subsurface environments. The framework provides a reproducible simulation-based foundation for subsequent validation using independent field or benchmark datasets.