HR: 0830h
AN: H11F-0909    [PDF]
TI: Evolution of Neural Networks for the Prediction of Hydraulic Conductivity as a Function of Borehole Geophysical Logs: Shobasama Site, Japan
AU: Reeves, P
EM: pcreeve@sandia.gov
AF: Sandia National Laboratories, PO Box 5800 MS 0735, Albuquerque, NM 87185-0735 United States
AU: * McKenna, S A
EM: samcken@sandia.gov
AF: Sandia National Laboratories, PO Box 5800 MS 0735, Albuquerque, NM 87185-0735 United States
AU: Takeuchi, S
EM: Takeuchi@tono.jnc.go.jp
AF: Japan Nuclear Cycle Development Institute, Tono Geoscience Center, 1-63, Yamanouchi, Akeyo, Mizunami, Gif 509-6132 Japan
AU: Saegusa, H
EM: saegusa@tono.jnc.go.jp
AF: Japan Nuclear Cycle Development Institute, Tono Geoscience Center, 1-63, Yamanouchi, Akeyo, Mizunami, Gif 509-6132 Japan
AB: In situ measurements of hydraulic conductivity in fractured rocks are expensive to acquire. Borehole geophysical measurements are relatively inexpensive to acquire but do not provide direct information on hydraulic conductivity. These geophysical measurements quantify properties of the rock that influence the hydraulic conductivity and it may be possible to employ a non-linear combination of these measurements to estimate hydraulic conductivity. Geophysical measurements collected in fractured granite at the Shobasama site in central Japan were used as the input to a feed-forward neural network. A simple genetic algorithm was used to simultaneously evolve the architecture and parameters of the neural network as well as determine an optimal subset of geophysical measurements for the prediction of hydraulic conductivity. The initial estimation procedure focused on predicting the class of the hydraulic conductivity, high, medium or low, from the geophysical measurements. This estimation was done while using the genetic algorithm to simultaneously determine the most important geophysical logs and optimize the architecture of the neural network. Results show that certain geophysical logs provide more information than others- most notably the short-normal resistivity, micro-resistivity, porosity and sonic logs provided the most information on hydraulic conductivity. The neural network produced excellent training results with accuracy of 90 percent or greater, but was unable to produce accurate predictions of the hydraulic conductivity class In the second phase of calculations, the selection of geophysical measurements is limited to only those that provide significant information. Additionally, this second phase predicts transmissivity instead of hydraulic conductivity in order to account for the differences in the length of the hydraulic test zones. Resulting predictions of transmissivity exhibit conditional bias with maximum prediction errors of three orders of magnitude occurring at the extreme measurement values. Results of these simulations indicate that the most informative geophysical measurements for the prediction of transmissivity are depth and sonic velocity. The long normal resistivity and self potential geophysical measurements are moderately informative. In addition, it was found that porosity and crack counts (clear, open, or hairline) do not inform predictions of transmissivity. This work was funded by the Japan Nuclear Cycle Development Institute. Sandia is a multiprogram laboratory operated by Sandia Corporation, a Lockheed Martin Company, for the United States Department of Energy's National Nuclear Security Administration under contract DE-AC04-94-AL-85000
DE: 1829 Groundwater hydrology
DE: 1894 Instruments and techniques
SC: Hydrology [H]
MN: 2003 Fall Meeting