Abstract: Permeability Prediction Using Artificial Neural Networks

J.H. Fang, H.C. Chen, David Kopaska-Merkel

Predicting permeability in uncored wells is one of the central problems in reservoir characterization. The prediction of permeability is very difficult for

two main reasons: (1) it is not yet possible to quantitatively describe the heterogeneity of porous media and (2) the flow field is the solution of a set of partial differential equations which are not easy to compute. One way to predict permeabilities of uncored wells is by porosity measurements from well logs. This is a viable approach because (1) porosity log measurements do not have to depend on expensive cores and (2) porosity and permeability are often correlated. Regression analysis is used to establish this linear or exponential relationship, and then to estimate permeabilities. However, this approach has not been very satisfactory because of poor agreement between observed and calculated permeability values.

We have employed a new approach, that of artificial neural networks, ANN, to do prediction. Unlike conventional statistical methods, this approach does not require sophisticated mathematics and tons of statistical data. Neural networks possess the ability to learn from examples. In our implementation, we first train the neural network on the porosity and permeability data. During the training phase, we employ incremental learning procedures designed for prediction. The trained network uses data from multiple cored wells to predict permeability variation in multiple non-cored wells. This permits maximum use of core data. We believe ANN is effective and efficacious in predicting permeabilities.

AAPG Search and Discovery Article #90986©1994 AAPG Annual Convention, Denver, Colorado, June 12-15, 1994