Geophysical subsurface characterization plays a key role in the success of geologic carbon sequestration (GCS). While deterministic inversion methods are commonly used due to their computational efficiency, they often fail to adequately quantify the model uncertainty, which is essential for informed decision-making and risk mitigation in GCS projects. In this study, we propose the SVGD-AE method, a novel geostatistical inversion approach that integrates geophysical data with prior geological knowledge to estimate subsurface properties. SVGD-AE combines Stein Variational Gradient Descent (SVGD) for sampling high-dimensional distributions with an autoencoder (AE) neural network for re-parameterizing reservoir models, aiming to accurately preserve geostatistical characteristics of reservoir models derived from geological priors. Through two synthetic examples, we demonstrate that the SVGD-AE method outperforms traditional probabilistic methods, particularly in inverse problems with complex posterior distributions. Then, we apply SVGD-AE to the Illinois Basin - Decatur Project (IBDP), a large-scale CO2 storage initiative in Decatur, Illinois, USA. The resulting petrophysical models with quantified uncertainty enhance our understanding of subsurface properties and have broad implications for the feasibility, decision making, and long-term safety of CO2 storage at the IBDP.

Geophysical monitoring of geologic carbon sequestration is critical for risk assessment during and after carbon dioxide (CO2) injection. Integration of multiple geophysical measurements is a promising approach to achieve high-resolution reservoir monitoring. However, joint inversion of large geophysical data is challenging due to high computational costs and difficulties in effectively incorporating measurements from different sources and with different resolutions. This study develops a differentiable physics model for large-scale joint inverse problems with reparameterization of model variables by deep neural networks and implementation of a differentiable programming approach of the forward model. The main novelty is the use of automatic differentiation and parallel computing for efficient multiphysics data assimilation. The application to the Sleipner benchmark model demonstrates that the proposed method is effective in estimation of reservoir properties from seismic and resistivity data and shows promising results for CO2 storage monitoring.