Uncertainty Quantification Based on Deep-Learning Approach Integrating Time-Lapse Seismic Data for Geological Carbon Storage

Abstract Carbon capture and storage (CCS) is a crucial technology for reducing greenhouse gas emissions to achieve net-zero goals by 2050. Reasonable assessment of CO2 plume behavior through reliable subsurface characterization and continuous monitoring (e.g., time-lapse seismic) is a prerequisite for the successful implementation CCS. However, the scarcity of data acquisition and the high degree of error during seismic inversion have hindered successful subsurface characterization and monitoring for CCS in many previous attempts. In this study, we propose a novel workflow that integrates time-lapse seismic data into subsurface model characterization with the assistance of deep learning. The suggested workflow demonstrates enhanced reservoir characterization performance and accurate prediction of future CO2 plume behavior. The study consists of three main components: (1) a seismic forward model, which generates synthetic time-lapse seismic data from relevant acoustic attributes such as porosity, density, and P-wave velocity; (2) a deep learning model based on generative adversarial networks (GANs), which inputs seismic data and outputs porosity and facies properties; and (3) a demonstration of the workflow in an anticline saline aquifer. By integrating initial and 5 years postinjection seismic data, the proposed workflow enables the creation of a more accurate ensemble of subsurface models compared to the initial ensemble. This approach effectively handles multiple possible geological scenarios and added noise in the seismic data, resulting in better predictions of future CO2 plume behavior.