Groundwater contaminant source identification via Bayesian model selection and uncertainty quantification

Many deterministic and stochastic approaches have been applied for groundwater contaminant source identification (GCSI) in recent decades. Usually, these implementations are based on a single groundwater model or fixed model structure and ignore the uncertainty of model structure. However, model structure uncertainty is inevitable for groundwater modeling, especially for complex geological environments and limited observations. This study evaluated the impact of model structure uncertainty on GCSI, and proposes an approach for GCSI based on Bayesian model selection. In the framework of multiple model analysis, a set of alternative model structures are used to represent the unknown groundwater system. Then, a novel nested sampling algorithm, POLYCHORD, is used for model selection and source identification. This algorithm is capable of estimating the model's marginal likelihood and inferring the posterior distribution of the contaminant source's characteristics simultaneously. Finally, this proposed approach is verified through two GCSI case studies, which include a synthetic groundwater contamination problem and a groundwater transport column experiment. The results demonstrated that GCSI could be inconsistent when using different model structures. Models with higher marginal likelihoods tend to have better performance on the predictions of the contaminant source's characteristics. It was concluded that POLYCHORD is efficient in marginal likelihood estimation and GCSI.