maria carrizo

Frequency-domain electromagnetic induction measurements are acquired to understand the distribution of electrical properties in the subsurface. Because the electromagnetic induction method is widely used in geophysical applications to reveal information about the subsurface, the algorithms for converting measurements into electrical conductivity distributions, or other petrophysical parameters are numerous. This geophysical method is widely applied since it is easily deployable and can cover large survey areas in a relatively small amount of time. A wide range of approaches, from simple linear relationships to sophisticated inversion methods, have been used to transform measurements into subsurface parameters. However, many of these methods depend on additional geological or geophysical information to constrain and validate the inversion result.

Near-surface applications of electromagnetic geophysical methods are crucial for society because they provide cost-effective non-invasive ways to investigate the subsurface for water management, natural resources exploration, and environmental hazards. Therefore, the appropriate use and accuracy of the estimations using this measurements is vital for resource assessment, effective hazard mitigation, and reliable engineering projects by providing a precise understanding of subsurface conditions. In general, this geophysical method has been studied exhaustively, but is still challenging to understand the uncertainties associated.

This thesis presents a comprehensive study of inversion methodologies for frequency domain electromagnetic induction surveys, addressing both computational efficiency and model reliability. We start by studying the sensitivity of the measurements to the electrical conductivity of the subsurface. We analyze the impact of using linear approximations and compare them to a global search strategy using precomputed lookup tables of semi-analytic forward responses. The global search method enables rapid electrical conductivity model estimation, without iterative forward modeling. This approach is validated with synthetic and field datasets, demonstrating fast convergence and reliable recovery of horizontally 2-layered earth models. Furthermore, the study demonstrates the added value of using both quadrature and in-phase components of the measurements. Numerical experiments demonstrate that incorporating the in-phase response enhances sensitivity to conductivity and, as a result, improves result accuracy compared with quadrature-only inversions, providing new insights into subsurface structure.

Moreover, this thesis investigates the well-posedness of frequency-domain electromagnetic induction data inversion. This analysis demonstrates that the uniqueness of one-dimensional subsurface models with two layers is ensured only when both quadrature and in-phase measurements are incorporated, for multi-coil single frequency configurations. On the contrary, for three-layer models the solution remains underdetermined without additional prior constraints. These findings highlight the fundamental limits of the inversion problem and the importance of exploiting the full information contained in the data.

Additionally, the limitations of piecewise inversion (subdividing the subsurface into one-dimensional datasets) and laterally constrained inversion are assessed under realistic three-dimensional conditions. Simulations reveal that lateral conductivity variations, topographic effects, and instrument tilt introduce significant distortions, even for small deviations from ideal assumptions. Numerical examples and field applications demonstrate that, while laterally constrained inversion provides smoother results than piecewise inversion, both methods remain vulnerable to error propagation in complex environments.

Finally, this thesis proposes a novel approach based on machine learning inversion via gradient boosted decision trees to estimate layered subsurface electrical conductivity distributions from the measurements. This prediction algorithm provides three clear advantages: Firstly, the ability of the decision trees to provide individual influence weights to different measurement parameters allows for a better understanding of the sensitivity of the individual data values to the subsurface properties. Secondly, this machine learning algorithm learns the physical system’s non-linearity by understanding how the different data parameters interact. Finally, predictions from the electrical conductivity model are generated without prior knowledge of the subsurface. Compared to optimization-based estimations, the machine learning resulting predictions more closely match the true models.

Together, these contributions advance the knowledge about the inversion of electromagnetic induction measurements by evaluating estimation methodologies, analyzing data utilization, and critically evaluating their application into electrical conductivity distribution of the subsurface. The results emphasize both the potential and the limitations of current inversion strategies, and they motivate the development of forward modelling and inversion methodologies for realistic frequency-domain electromagnetic induction survey conditions. The main contribution of this study is an easily applicable machine learning estimation method without additional information of the subsurface.