Permeability Predictions From Dielectric Dispersion Logs Using Supervised Machine Learning: A Johan Sverdrup Field Example

Abstract

Permeability models largely rely on core measurements as input. The propagation of these models beyond the cored interval is often by use of the empirical porosity-permeability relationship. The problem is that porosity itself does not contain information about the pore geometry which controls permeability, hence these relationships carry high uncertainty in uncored intervals and nearby wells. Dielectric dispersion, on the other hand, is inherently linked to the pore geometry since it is sensitive to charge build up at the rock-fluid interfaces of the interconnected pore network through the Maxwell-Wagner effect. We aim to utilize this connection between pore geometry and dielectric dispersion to predict permeability using a core-data trained supervised machine learning model on dielectric dispersion wireline logging arrays. It builds upon a previous single-well study (Norbisrath, 2018) where the main concern was the repeatability in other wells, which is now addressed here. The study area is the Johan Sverdrup field on the Norwegian Continental Shelf. Data consists of core plug permeabilities and dielectric dispersion wireline logs from five wells. Capturing the dielectric frequency dispersion involves determining the slope of both attenuation and phase shift measurements made at different frequencies and transmitter-receiver spacings (feature engineering). The model will be trained on a subset of the core data (supervised machine learning), and subsequently propagated along the entire logged interval, as well as to the test well which was not part of the training set. Hyperparameter tuning will be used to optimize the model, and cross-validation used to prevent overfitting. Preliminary results show that the dielectric dispersion logging data contains enough information about the pore geometry to accurately describe and predict core plug permeabilities, not only in the same well but also in nearby wells that were not used in training the model. Correlation coefficients between estimated and predicted core permeability values are around R = 0.8. Given additional training input data and ground truthing in other wells, the described method could potentially reduce the need for coring when dielectric dispersion wireline logs are run. In the future we aim to explore the possibility of using dielectric dispersion data from LWD (Logging While Drilling) resistivity propagation tools as input for our permeability predictions. This would greatly enhance formation evaluation since these data are readily available in thousands of wells and are generally acquired in every new well. A model trained on a large amount of existing core data could enable real time permeability predictions from LWD tools.