Abstract We rely on a global sensitivity analysis (GSA) approach to identify the dominant physical and biogeochemical controls on dissolved oxygen (DO) dynamics in riparian aquifers. The study is motivated by the observation that availability of DO is key to regulating redox conditions and associated processes in the subsurface. Yet, the complexity of coupled flow and transport models, combined with model input uncertainty challenges our ability to fully characterize system behavior. To address this issue, we integrate Bayesian network‐based and variance‐based methods into a comprehensive GSA framework, enabling a robust evaluation of parameter and process sensitivities. To overcome the high computational demand of GSA for complex numerical models, we develop surrogate models using deep learning approaches (i.e., multi‐layer perceptrons and convolutional neural networks). Application of this framework to a high‐resolution model of riparian DO transport reveals that river stage dynamics (i.e., period and amplitude of water level fluctuations) are primary drivers of DO supply to the aquifer system. Hydraulic conductivity, riverine DO concentration, and the maximum DO reaction rate exhibit important but localized effects, influencing different transport pathways including river water infiltration, entrapped air dissolution, and diffusion through the unsaturated zone. In contrast, parameters such as porosity, longitudinal dispersion, and van Genuchten soil parameters exhibit negligible influence. These findings underscore the value of combining deep learning and GSA to efficiently evaluate complex environmental systems and to guide model simplification and diagnosis.
Dai et al. (Fri,) studied this question.