Abstract: Neural Network Interpretation of Well Log Data
LIN ZHANG and MARY POULTON, University of Arizona, Dept. of Mining and Geological of Engineering, Tucson, AZ
Neural networks are computer models loosely based on the function of biological neurons. As such, they find application in areas of pattern recognition where humans perform very well and traditional computer models perform poorly. One of the strengths of the neural network approach is that it can perform non-parametric modeling very rapidly and with a high degree of accuracy. In geophysics, neural networks have been used to invert data, classify patterns, and combine data from multiple sensors. Our application involves the use of neural networks to pick lithologic layers from wireline log data, to invert shallow EM log data to produce a model of resistivity and thickness, and to learn synthetic two-dimensional forward models of electric logs.
Our first step in processing log data is to break the log into sections based on lithologic layers. We have developed an adaptive layer picking code using neural networks that can accurately detect layers in log data. Next, we use a modular neural network architecture to invert data from a shallow EM induction tool (Geonics EM39) commonly used for groundwater studies. Conductivity information from the tool is converted to logl0 resistivity and used as input to a neural network, which in turn produces a model of resistivity and thickness for each section of the log. In another application, we use the neural network to learn the forward model response of various electrical logging tools. The trained networks can then be used in conjunction with an inversion routine to greatly reduce the amount of time it takes to perform a two-dimensional interpretation of log data.
AAPG Search and Discovery Article #90931©1998 AAPG Foundation Grants-in-Aid