NEURAL NETWORK PREDICTION OF PERMEABILITY IN THE EL GARIA FORMATION, ASHTART OILFIELD, OFFSHORE TUNISIA
J.H. Ligtenberg* and A.G. Wansink*
*deGroot-Bril Earth Sciences, Boulevard-1945 nr. 24, 7511 AE, Enschede, The Netherlands.
Corresponding author: J.H. Ligtenberg, email herald@dgb.nl
The Lower Eocene El Garia Formation forms the reservoir rock at the Ashtart oilfield, offshore Tunisia. It comprises a thick package of mainly nummulitic packstones and grainstones with variable reservoir quality. Although porosity is moderate to high, permeability is often poor to fair with some high permeability streaks.
The aim of this study was to establish relationships between log-derived data and core data, and to apply these relationships in a predictive sense to uncored intervals. An initial objective was to predict from measured logs and core data the limestone depositional texture (as indicated by the Dunham classification), as well as porosity and permeability. A total of nine wells with complete logging suites, multiple cored intervals with core plug measurements together with detailed core interpretations were available.
We used a fully-connected Multi-Layer-Perceptron network (a type of neural network) to establish possible non-linear relationships. Detailed analyses revealed that no relationship exists between log response and limestone texture (Dunham class). The initial idea to predict Dunham class, and subsequently to use the classification results to predict permeability, could not therefore be pursued. However, further analyses revealed that it was feasible to predict permeability without using the depositional fabric, but using a combination of wireline logs and measured core porosity. Careful preparation of the training set for the neural network proved to be very important. Early experiments showed that low to fair permeability (1-35 mD) could be predicted with confidence, but that the network failed to predict the high permeability streaks. "Balancing" the data set solved this problem. Balancing is a technique in which the training set is increased by adding more examples to the under-sampled part of the data space. Examples are created by random selection from the training set and white noise is added. After balancing, the neural network’s performance improved significantly. Testing the neural network on two wells indicated that this method is capable of predicting the entire range of permeability with confidence.