A parallel workflow framework using encoder-decoder LSTMs for uncertainty quantification in contaminant source identification in groundwater

Abstract

Groundwater contamination is a major problem and it can be contaminated by natural or manmade activities. Contaminant sources cannot be identified promptly after a contamination event due to slow groundwater movement in aquifers. To limit the further spread of the plume and to take appropriate remedial measures, it is essential to determine the release rates associated with source locations as soon as a contaminant is detected in one of the observation wells in the aquifer. Conventional techniques for contaminant source identification such as the simulation–optimization (SO) uses a simulation model, which is based on estimated or, measured hydrogeological parameters, for replicating the aquifer response to contamination at source locations. However, the associated intrinsic uncertainty in the parameters can affect the estimated release histories at the source locations. To assess confidence in the results, it is important to consider stochasticity in the hydrogeological parameters for contaminant release history estimation requiring numerous runs of the SO model for corresponding breakthrough curves. The present study proposes a novel surrogate simulation optimization (SSO) framework which utilizes an Encoder-Decoder Long Short-Term Memory (ED-LSTM) based surrogate model along with Multiverse Optimization (MVO) for contaminant source identification. The comparison of ED-LSTM with other surrogate models shows its better accuracy and significantly low computational time requirement compared to the simulation model. To enhance the computational efficiency further, the SSO model is embedded in a parallel workflow for the quantification of uncertainty associated with the inverse modelling. The study shows that the proposed workflow can be efficiently used for uncertainty quantification in contaminant source identification due to stochastic parameters with a reduced computational cost.