Uncertainty analysis of contaminant sources in fracture networks

Abstract

Identifying contaminant source characteristics is essential for groundwater remediation, especially in fracture networks where source uncertainty complicates assessment. Traditional inverse methods often require extensive forward simulations and rely on explicit likelihood and error model specifications, which can be computationally demanding and challenging under uncertainty. Here, we propose to quantify contaminant source uncertainty in fracture networks using Bayesian Evidential Learning (BEL) that learns the relationship between observed breakthrough curves (BTCs) and source characteristics from an offline training ensemble, thereby reducing computational burden and mitigating sensitivity to subjective likelihood specifications. Training data consist of the targets (contaminant source location, release time and concentration), sampled from their prior distribution, and the corresponding predictors (BTCs and their statistical features) obtained by forward simulations in a fracture network with hydrogeological uncertainties (inflow velocity and fracture aperture). The Robust Mahalanobis Distance (RMD) was applied to multidimensional outlier detection, falsifying source locations inconsistent with observations. Consistent source locations were then discretized using one-hot encoding. Principal component analysis (PCA) and canonical correlation analysis (CCA) were employed to establish joint probability distribution functions between predictor and target. We then applied the learned relationship on laboratory and synthetic data of solute transport in fracture networks to predict the posterior source distributions. The BEL posterior guides a brute force Monte Carlo random search that refines contaminant source parameters by minimizing the misfit between simulated and observed BTC, improving identifiability. Results accurately predict experimental values, effectively quantifying contaminant source uncertainty in fracture networks and providing a novel approach for tracking groundwater contamination.