Non-gaussian hydraulic conductivity and potential contaminant source identification: A comparison of two advanced DLPM-based inversion framework

Abstract

Accurate identification of hydraulic conductivity fields (K) and contaminant source parameters is imperative for the enhanced assessment and effective remediation of polluted aquifers. Given the challenges posed by non-Gaussian distributions, high dimensionality, and ill-posed nature of groundwater inversion problems, reducing unknowns is a common strategy. Unlike conventional parameterization methods constrained by prior assumptions, this research introduces an innovative deep learning-based parameterization method (DLPM), AEdiffusion. AEdiffusion combines a Diffusion Denoising Probabilistic Model (DDPM) with a Variational Autoencoder (VAE) through a generator-refiner strategy, enabling the generation of high-dimensional K fields from low-dimensional latent representations. Additionally, this study examines the application of Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP), another advanced DLPM, in groundwater inversion. Through comparative analysis within a Data Assimilation (DA) framework, focusing on non-Gaussian K fields and three potential contaminant sources under varied data availability scenarios, this study reveals that both DLPM-based inversion frameworks are capable of identifying K fields and the true contaminant sources. Notably, the AEdiffusion-based framework excels in extracting critical information from sparse observations, delivering more stable performance, but at the cost of increased time consumption compared to WGAN-GP-based framework.