Quantifying Groundwater Contaminant Source Uncertainty in Fracture Networks Combining Falsification and Bayesian Evidential Learning

Identifying contaminant sources is crucial for managing groundwater contamination, particularly in complex fracture networks. Traditional methods for source identification often face limitations such as sensitivity to data perturbations, reliance on simplified hydrological models, and challenges in handling the ill-posed nature of the inverse problem. This study introduces a novel application of Bayesian Evidential Learning (BEL) to quantify contaminant source uncertainty in fracture networks(Hermans et al., 2018; Thibaut et al., 2021).

BEL relies on learning a direct relationship between the target parameters (source location, release time, and concentration) and predictors (breakthrough curves (BTCs) and their statistical features). The learning step relies on the sampling of target parameters for which the release and transport of contaminant is simulated, and the resulting BTCs at the observation point extracted. The complexity of the training process was mitigated by incorporating falsification to classify the prior model(Yin et al., 2020). One-hot encoding then was employed to discretize potential source locations, enhancing the correlation between predictor and target using principal component analysis (PCA) and canonical correlation analysis (CCA) (Figure 1). Experimental data and numerical simulations of solute transport in fracture networks were then employed to validate the BEL framework (Figure 2).

Results demonstrate that BEL not only achieves accurate predictions on the source location, release time and concentration, but also provides robust uncertainty quantification for contaminant sources. These findings highlight BEL's potential as a powerful tool for improving source tracking and remediation strategies in groundwater systems. Future research should consider uncertainty in the fracture network and hydraulic properties of the fractures.

Figure 1. Multivariate analysis of training data. A. The explanatory power of data across different PCs in the PCA space. B-E are the bivariate distributions of predictor and target data in the CCA space.

Figure 2. Posterior distribution predictions for contaminant source information. Red lines correspond to test data

 

References

Hermans, T., Nguyen, F., Klepikova, M., Dassargues, A., & Caers, J. (2018). Uncertainty Quantification of Medium-Term Heat Storage From Short-Term Geophysical Experiments Using Bayesian Evidential Learning. Water Resources Research, 54(4), 2931–2948. https://doi.org/10.1002/2017WR022135

Thibaut, R., Laloy, E., & Hermans, T. (2021). A new framework for experimental design using Bayesian Evidential Learning: The case of wellhead protection area. Journal of Hydrology, 603, 126903. https://doi.org/10.1016/j.jhydrol.2021.126903

Yin, Z., Strebelle, S., & Caers, J. (2020). Automated Monte Carlo-based quantification and updating of geological uncertainty with borehole data (AutoBEL v1.0). Geoscientific Model Development, 13(2), 651–672. https://doi.org/10.5194/gmd-13-651-2020