Three-axis borehole gravity monitoring for CO2 storage using machine learning coupled to fluid flow simulator

Abstract

The field of geophysics faces the daunting task of monitoring complex reservoir dynamics and imaging carbon dioxide storage up to several decades into the future. This presents numerous challenges, including sensitivity to parameter changes, resolution of obtained results and the cost of long-term deployment. To effectively store CO₂ subsurface, it is necessary to monitor and account for the injected CO₂. The gravity method provides several advantages for CO₂ monitoring, as changes in fluid saturation correspond directly and uniquely to observed density changes. Three-axis borehole gravity has demonstrated significant promise as a next-generation tool for reliably monitoring reservoir dynamics across a range of depths and sizes. However, the gravity inverse problem is highly ill-posed, necessitating regularization that incorporates prior knowledge. To address this issue, we propose using a feed-forward neural network, a machine learning method, to invert time-lapse three-axis borehole gravity data and monitor CO₂ movement within a reservoir. By training the neural network on models that analyse changes in density and corresponding gravity responses resulting from perturbations made to the reservoir model, we can create scenarios that train the algorithm to identify unexpected CO₂ migration in addition to the normal movement of CO₂. Our method is demonstrated using reservoir models for the Johansen formation in offshore Norway. We convert reservoir saturation models into density changes and generate their corresponding three-axis gravity data in a set of boreholes. Our results show that the developed machine learning inversion algorithm has high reliability and resolution for imaging density change associated with CO₂ plumes, as demonstrated in the Johansen reservoir models utilized by the simulator. We also investigate machine learning inversion using regularization parameters and show that it is robust, with a strong tolerance for higher levels of noise. Our study demonstrates that the developed machine learning algorithm is a powerful tool for inverting three-axis borehole gravity data and monitoring the migration and long-term storage of injected CO₂.