 |
|
Kansas Geological Survey Open-file Report 2003-68 |
The use of artificial intelligence tools such as neural networks, fuzzy logic or other classification tools to help solve geologic problems has blossomed in the past fifteen years. A histogram of the number of literature “hits” on the key words "neural networks" using the GeoRef search engine illustrates that point.
For lithofacies prediction in this project we have chosen to use a standard, public domain, single hidden-layer neural network. Input-layer nodes are the selected geophysical well logs (e-log curves) and geologic constraining variables, and output-layer nodes are the set of lithofacies membership probabilities. For ease of use, Bohling added Visual Basic code for neural network training and prediction to Kipling.xla, an existing Excel add-in for nonparametric regression and classification that was developed by Bohling and Doveton at the KGS. Code for the batch application of the trained neural network to predict lithofacies for a large set of well files was also implemented in the Excel add-in, providing a means for computing lithofacies membership probabilities at hundreds of wells in a matter of minutes.
Predictor variables were chosen by trial and error. Neural network parameters
such as the number of hidden-layer nodes, which governs the richness of the
model, and the damping parameter, which constrains the magnitude of the network
weights to help prevent over training, were optimized by automated cross-validation
with keystone well data.
    |
e-mail : webadmin@kgs.ku.edu
Last updated January 2004
http://www.kgs.ku.edu/PRS/publication/2003/ofr2003-68/slide11.html