Predicting Permeability in Real-Time from LWD Resistivity and Gamma Ray Logs

Abstract Real-time permeability prediction and flow risk assessment is useful both for the safety of drilling operations as well as operational challenges like determination of pressure point depths and perforation zones. A model giving accurate permeability predictions based solely on LWD (logging-while-drilling) data can potentially save days of rig time by eliminating the need for wireline or coring runs. Drilling safety profits from early detection of flow zones that can destabilize the wellbore and require timely treatment. Our goal is to implement these predictive models into real-time geo-operational drilling software which can then be used by geoscientists and engineers alike. The novelty of this study is the large amount of available training data. We build upon ongoing efforts to lift both legacy and newly acquired petrophysical log and core data to the cloud, creating a training dataset on a continent-wide scale. The core data contains about 650,000 core permeability measurements, around 43,000 of which have corresponding (on-depth) logging-while-drilling (LWD) resistivity propagation and gamma ray curves. The combined dataset is mainly based on wells from the NCS (Norwegian Continental Shelf). This extensive training dataset is vital for developing universally applicable models across different formations of the various basins on the NCS. Nonetheless, since the data is from various vintages and vendors, a large part of the work evolves around data preparation, cleaning, and normalization. Standard machine learning techniques are then used to train the model and assess its predictability on new data from unseen wells. Results show that a model that uses just LWD gamma ray and resistivity curves as input can predict unseen data in new wells with a correlation coefficient of r = 0.70. Since the input logs are ubiquitous in basically any hole size, the model can be used to both identify shallow hazards (flow zones) as well as permeable target intervals in the reservoir. Adding in neutron and density curves, typical for reservoir section drilling, the model’s accuracy increases to r = 0.85.