Improving deep groundwater aquifer characterization with deep learning inversion of audio-frequency magnetotelluric data

Abstract

Deep groundwater is a crucial resource for drinking, industry, and ecosystems. However, its extensive subsurface distribution poses challenges for traditional hydrological sampling methods. Audio-frequency Magnetotelluric (AMT) is commonly used to image shallow subsurface resistivity distribution, but its application for quantifying deep groundwater aquifers is challenging. In this study, we propose a two-step strategy deep learning (DL) to estimate subsurface water information from AMT data. First, we employ an inversion DL network to directly predict subsurface resistivity distribution from AMT data. To assess the impact of data and prior information on DL inversions, we progressively include more input data during training, from apparent resistivity to apparent resistivity, phase, and Bostick transformed initial models. In the second step, we use another DL network to predict subsurface structure from AMT inversion results. We then estimate water content based on unit-specific petrophysical relationships. Evaluating our algorithm with synthetic cases, we find that including more information generally improves network performance, particularly when incorporating initial Bostick transformed models. Applying the algorithm to a field hydrogeological survey, we compare the inversion network trained with apparent resistivity, phase, and Bostick transformed initial models to traditional regularization inversions. The new approach shows better consistency with borehole data. Another network extracts structure information, enabling the estimation of subsurface water content and gaining valuable hydrogeological insights. To summarize, our study presents a novel DL-based approach to quantitatively delineate deep groundwater aquifers using AMT data. The proposed algorithm demonstrates promising performance in estimating subsurface hydrologic properties, providing valuable insights for hydrogeological investigations.