Groundwater pollution source identification using a hybrid Metropolis-Hastings and particle filter algorithm

Groundwater pollution source identification (GPSI) plays a vital role in groundwater protection and remediation. Among various inversion approaches, the Metropolis-Hastings (MH) and particle filter (PF) algorithms have been widely employed for probabilistic parameter estimation. However, the MH algorithm often suffers from slow convergence and a tendency to become trapped in local optima, while the PF algorithm experiences particle degeneracy and high computational cost in high-dimensional nonlinear systems. To address these challenges, this study proposes a hybrid Metropolis-Hastings and particle filter (PF-MH) algorithm that integrates the global exploration of PF with the local optimization of MH, effectively improving sampling efficiency and stability in posterior estimation. Furthermore, a residual neural network (ResNet) surrogate model is incorporated to approximate the numerical simulator, greatly reducing computational burden without compromising accuracy. We designed two hypothetical aquifer cases of varying complexity to validate the proposed method's effectiveness. The ResNet surrogate model significantly outperformed the multilayer perceptron (MLP) in both cases. Moreover, the PF-MH algorithm outperformed the MH algorithm, reducing average relative errors in pollutant release and hydrogeological parameter estimates while demonstrating greater stability across multiple runs. Overall, integrating the ResNet surrogate model with the PF-MH algorithm provides an efficient and reliable approach to tackle the high computational costs and significant uncertainties in GPSI.