Full text loading...
-
Variational Inference Inversion of Geoelectrical Resistivity Data
- Publisher: European Association of Geoscientists & Engineers
- Source: Conference Proceedings, NSG 2025: 6th Conference on Geophysics for Mineral Exploration and Mining, Sep 2025, Volume 2025, p.1 - 5
Abstract
Summary
Electrical Resistivity Tomography (ERT) is a popular geophysical technique for characterizing subsurface resistivity distributions. However, ERT inversion is an inherently ill-posed problem and traditional deterministic inversion techniques are prone to local minima and strongly depend on the initialization. Moreover, they are not able to quantify uncertainty, crucial for decision-making in applied geophysics. In this study, we present a computationally efficient Bayesian framework for ERT inversion, based on Annealed Stein Variational Gradient Descent (A-SVGD) and a Discrete Cosine Transform (DCT) reparameterization of the model space. This approach enables uncertainty quantification while reducing the dimensionality of the inverse problem, thus maintaining tractable computational costs. We apply our method to a real dipole-dipole ERT dataset and compare the results with those from a classical deterministic inversion. Our approach yields a significantly lower data misfit and produces a resistivity model consistent with the deterministic solution. Moreover, the final particle ensemble allows quantification of model uncertainty. The DCT compression is crucial to maintain affordable computational costs, at the expense of a slightly lower model resolution. Our results highlight the potential of A-SVGD combined with DCT compression as a powerful alternative to conventional inversion techniques, also able to provide accurate uncertainty quantification maintaining computational efficiency.
© EAGE Publications BV