| Abstract |
Finding naturally occurring porous and permeable reservoir rocks that allow sufficient fluid flow for economical geothermal applications relies on the use of geophysical measurements sensitive to the rock composition such as seismic amplitudes, seismic velocities, electromagnetics-derived resistivity or gravity-derived density. Tight budgets usually means that the quality and coverage of these data is suboptimal in addition to intrinsic resolution limitations for each technique. To partly overcome these limitations, joint inversions have been proposed to combine their strengths while alleviating their individual weaknesses. Although theoretically appealing, access and ease of use have limited their practical use. But recent advances in open-access user-friendly machine learning algorithms have been a game-changer. Building on existing work, we demonstrate how a series of UNet-based neural networks can be used to estimate porosity from multiphysics data. The innovative aspect of this work relies on the treatment of the input and target data provided to the networks. As the mentioned geophysical measurements have different frequency bandwidths, each input data is first spectrally decomposed in a series of frequency bands. Networks with frequency-dependent parameters are then used to estimate the porosity in each band based on the data available in each frequency range. The estimates from all networks are then recomposed in the frequency domain to produce a porosity estimate in the spatial domain. The proposed workflow is first tested on a synthetic dataset to evaluate its predictive capability and then on a real dataset where the estimated porosity is compared to wellbore-derived porosity. |