Estimating line contaminant sources in non-Gaussian groundwater conductivity fields using deep learning-based framework

This study applies two hybrid inversion frameworks, GAN-ILUES and GAN-OANW-ILUES, to accurately estimate the line contaminant source information and hydraulic conductivity in non-Gaussian groundwater fields. These frameworks combine data assimilation and deep learning methods, using WGAN-GP as a generator model for non-Gaussian hydraulic conductivity field and a CNN-based surrogate model, OANW, for surrogating the original forward groundwater model. The effectiveness and feasibility of WGAN-GP and OANW are verified separately. Then, the frameworks are verified on a 64 m by 64 m channelized aquifer with a linear contaminant source using measurement data from 20 observation wells. Results show that both frameworks can estimate model parameters and reproduce the contaminant concentration field. Compared with GAN-ILUES, GAN-OANW-ILUES with OANW surrogate model significantly improves simulation efficiency at a price of slightly larger model deviations. The inversion time for GAN-OANW-ILUES is less than 5% of that for GAN-ILUES. This study demonstrates the potential of combining data assimilation and deep learning methods for improving the performance of traditional inversion methods on tracing line contaminants in non-Gaussian fields. •Two deep-learning methods are integrated with data assimilation for inverse modeling of groundwater pollution.•The proposed modeling framework successfully traces the line contaminant source in non-Gaussian conductivity fields.•Model error slightly increases when the conductivity changes drastically.•The combined inversion framework significantly enhances inversion efficiency.