Machine learning inversion of time-lapse three-axis borehole gravity data for CO2 monitoring

Abstract

Effective subsurface CO2 storage necessitates monitoring and accounting for injected CO2. Gravity methods are an enticing appliance for CO2 monitoring since the change in density observed by the gravity method is directly related to fluid saturation. In particular, three-axis borehole gravity has shown to be a next-generation tool taking us into the future for reliably monitoring reservoir dynamics in fields with a wide range of depths and sizes. Here, we use a machine learning (ML) method, namely the feed-forward neural network, to invert time-lapse three-axis borehole gravity data for monitoring the movement of injected CO2 within a reservoir. The neural network is trained using change in density and the associated gravity responses derived from perturbations to a field’s reservoir model. In addition to normal migration, these perturbations additionally include scenarios that can be used to train the algorithm to detect unexpected leakage of CO2 beyond the field. The method is developed and demonstrated here using a fluid flow reservoir simulator for the Johansen formation, offshore Norway. Results show that the reconstruction capabilities of the ML inversion are highly reliable, with resolution similar to the Johansen reservoir models utilized by the simulator.