Carnarvon Basin: Mungaroo Formation, The Challenges in EstimatingPermeability

摘要

There was always the challenge to match Permeability estimated from logs to Permeability derived fromwell tests. The main reason for this is the difference in scale, both vertically in the borehole (net contributingsands) as well as radially out into the formation. The well test that takes into consideration many hours offluid flow into the borehole was always deemed more representative than permeability derived from logs thatmeasures a smaller area and volume. However, this perception should not relegate log-derived permeabilityto an insignificant parameter within dynamic models. When wells are stimulated prior to well test (e.g.under-balanced perforation or chemical stimulation for clean-up), it is expected that the permeability fromthe well test be enhanced as seen in Iago-2 and Gorgon-3 wells. This paper takes core data from wells in the Carnarvon Basin, create log relationships to predictpermeability that match these core data, and compare these to wireline formation tester mobility and to welltests. The results are very promising, and the workflow proposed can be applied to any well in the basin.One of the objectives of this paper is to create a workflow that can be replicated easily and to use raw logsthat are available across all wells, in order to reduce the uncertainty in the predicted permeability. The reservoir sands of the Mungaroo formation are easily recognised by the cross over in the neutron-density logs, in both gas zones and water zones. This is the criteria used by operators to obtain formationpressure tests (MDT/RFT) and this is the same criteria used in this paper to define reservoir sands. Onlythose core data acquired in reservoir sands are used as the "Learning" dataset to predict permeability. Severallearning datasets were created, and these were blind tested on other wells that were not part of the learningdataset. The results of these predicted permeabilities were cross plotted against core permeability that havebeen over-burden corrected and depth shifted to wireline logs. Where the match is not satisfactory, newlearning datasets are derived and this step of the workflow is repeated. At the end, there are four groups oflearning datasets that are used as the final results.