Novel approaches in electromagnetic methods using reparameterization and machine learning with application to monitoring geological CO₂ storage sites

Abstract

Generalized geophysical inversion is a powerful tool to construct physical property models of the subsurface from measurements of different fields, such as electromagnetic or potential fields, at receiver locations. Geophysical inversion methods are computationally expensive in terms of memory, storage, and computational time. These limitations make inversion an impractical method to enhance geophysical surveys in real time, such as in determining where future receivers should be located during a field campaign, or when an answer is needed quickly, as in problems related to decisions about well operations. Within this thesis, I make use of unique model parameterizations and data characteristics to limit the computational cost and perform a rapid inversion of electromagnetics data for subsurface properties. The method of zero-level curves aids in determining the location and primitive shape of a compact conductor in the subsurface, and can be used to make operation decisions in real time. The zero-level curve can also be used to approximate a resistive body in a conductive media, such as a CO_2 plume in a saline reservoir. I then focus on the feasibility of monitoring supercritical CO_2 that has been injected into a conductive reservoir for a fictitious scenario based on the Kemper CarbonSAFE site in Mississippi. In CO_2 monitoring there is a push for low-cost geophysical methods to be part of the larger monitoring plan, making electromagnetics an attractive choice. Furthermore, focusing on the magnetic field from such surveys offers additional flexibility to monitoring due to the advent of drone-based sensors. I model electric dipole sources in the frequency domain, a time-domain current loop, and magnetotellurics to determine if the current generation of sensor technology is sensitive enough to measure the resulting secondary magnetic fields. I find that electric dipole sources and magnetotellurics produce measurable fields, but a time-domain current loop does not for this particular site. Finally, I apply a convolutional neural network to automatically invert magnetotelluric tipper data for a parameterized model of the CO_2 plume in a conductive reservoir. Parameterizing the model of the CO_2 allows the network to train faster when compared to methods that solve for a full conductivity model of the subsurface. The network is also applied in classification mode to determine if the CO_2 is contained in the reservoir layer, or if it has escaped the seal, however, the network fails to distinguish between these cases, likely due to the limited vertical resolution offered by magnetotelluric tipper data.