Groundwater contamination source estimation based on a refined particle filter associated with a deep residual neural network surrogate

Groundwater contamination source estimation (GCSE) involves an inverse process to match time-series monitoring data in sparse observation wells. It is commonly accompanied by a search task in high-dimensional space and huge computational burden brought about by massive callings of the simulation model. Particle filters can provide accurate estimation for a high-dimensional search task in source estimation, but the process suffers from particle degradation and huge computational load brought about by repeatedly solving the transport simulation model. To tackle the particle degradation, an iterative ensemble smoother was introduced to provide a proper proposal distribution, improving the search efficiency of the traditional particle filter. Moreover, to relieve the computational burden, a deep residual neural network was proposed to perform the surrogate task for the highly nonlinear and long-running-time original simulation model. In general, a refined particle filter with a deep-learning-method surrogate was proposed as an inverse framework for GCSE, which was evaluated by estimation tasks for a point-source contamination case and an areal-source contamination case, respectively, under different levels of observation errors. The results indicated that the deep-residual-neural-network surrogate model achieved the performance R-2 of 0.993 and 0.995, respectively for point-source and aerial-source contamination, to substitute the simulation models with a swift invoking process. Furthermore, the iterative ensemble smoother evidently improved the estimation efficiency of the particle filter. The proposed inverse framework can provide reliable and stable estimation of the groundwater contamination source and aquifer hydraulic conductivity.