An Integrated Application of Neural Network and Markov Chain Techniques to Prediction of Lithofacies from Well Logs

The Permian Council Grove Group in the Panoma Field of southwest Kansas has yielded 80 x 10 meter of gas from approximately 2600 wells from a 60-meter interval at depths of 800-1,000 meters since its discovery in the 1960’s. Initial gas saturation, production rates and cumulative production in the Panoma Field are controlled by the distribution of porosity and permeability in the field, which are in turn controlled by the distribution of facies. We have used a single hidden-layer neural network to compute facies membership probabilities from geophysical well logs measured in approximately 470 wells throughout the field. The network was trained using facies assignments from detailed core descriptions in eight wells. For ease of use, the neural net prediction (feed-forward) code has been implemented as part of an Excel add-in using Visual Basic. However, the training of the network via backpropagation is too computationally intensive a task for this environment, and so is accomplished through an automated invocation of the neural network function provided as part of the public-domain R language. Code for the batch application of the trained neural network to log data from a large number of LAS (Log ASCII Standard) files is also implemented in the Excel add-in, providing an easy means for computing facies membership probabilities at a large number of wells. We then use public-domain Markov chain simulation code to produce a gridded realization of the facies architecture throughout the field, conditioned on the facies probabilities computed by the neural network. The simulation employs a transition probability model based on the facies sequences observed in core for the eight training wells, augmented by geological understanding of the expected facies relationships. This process provides an ideal means for merging geological knowledge with the dense quantitative information provided by geophysical well logs.