Joint Identification of Groundwater Contamination Source and Heterogeneous Hydrogeological Parameters in LNAPL‐Contaminated Sites Based on Deep Convolutional Encoder‐Decoder Neural Networks

Abstract

Accurate identification of nonaqueous phase liquid (NAPL) contamination sources is critical in pollution management. However, the joint inversion of source characteristics and heterogeneous hydrogeological parameters faces two major challenges. First, the high dimensionality of the heterogeneous parameter field increases inversion complexity. Second, the prohibitive computational cost of multiphase flow simulation renders traditional numerical model‐based approaches infeasible. To overcome these issues, this study proposes an inversion framework integrating a deep convolutional encoder‐decoder neural network (DCEDN) as a surrogate model. The DCEDN excels at capturing high‐dimensional nonlinear relationships via its convolutional encoder‐decoder architecture, enabling efficient feature extraction and dimensionality reduction and making it suitable for this task. Two specific DCEDN variants‐Deep Dense Convolutional Network (DDCN) and Deep Residual Dense Convolutional Network (DRDCN) ‐ were developed as surrogate models and coupled with the iterative local updating ensemble smoother (ILUES) inversion algorithm. The proposed framework is applied to a LNAPL‐contaminated site in Jilin City, China, extending existing research that is often limited to solute transport in synthetic cases. The inversion yielded well‐converged posterior distributions for the contamination source and other hydrogeological parameters. Forward simulations using the joint maximum a posteriori (MAP) estimates successfully captured the overall pattern of the observed contaminant plumes ( R ²  = 0.9), indicating that the inferred parameter set reproduces the observed concentration data well. The computational cost of the entire inversion process was reduced to 1.92% of that required by a traditional, numerically intensive approach. This demonstrates the DCEDN‐based framework's effectiveness in enhancing the accuracy and efficiency of the joint inversion.