Abstract

Accurate estimation of reservoir permeability is vital in the design and implementation of a CO₂ flood. The best method for determining reservoir permeability is to model core-derived permeability data. However, most Permian Basin oil fields lack sufficient core coverage for core-based models. Therefore, the common method has been to develop linear relationships between core-derived porosity and permeability, then apply these relationships to porosity logs from non-cored wells. This method has limitations because the linear relationships are often poor.

Neural network technology provides an alternative method for determining reservoir permeability. Neural networks estimate permeability by learning the relationships between many reservoir characteristics, not just porosity.

Data from six cored wells in the San Andres reservoir of the Roberts Unit were loaded into a neural network designed to predict permeabilities. A backpropagation neural network with one hidden layer containing 30 processing elements was used. The network learned those relationships in 3.1 million iterations using as inputs: the geographic location of the cored well, subsea depths, core porosities and zones, and as the output the difference between core-derived permeabilities and linear regression-derived permeabilities. A correlation coefficient of 0.81 was calculated for the neural network-derived permeability values. This compares to a 0.44 correlation coefficient for the linear regression-derived permeability values.