AdvisorPoulton, Mary M.
MetadataShow full item record
PublisherThe University of Arizona.
RightsCopyright © is held by the author. Digital access to this material is made possible by the University Libraries, University of Arizona. Further transmission, reproduction or presentation (such as public display or performance) of protected items is prohibited except with permission of the author.
AbstractForward modeling and inversion plays a fundamental role in well-logging data processing and interpretation. Conventional modeling and inversion methods, however, are usually computer intensive and are not suitable for well-site applications. Neural networks provide a means for building fast and robust modeling and inversion algorithms that can be applied at well sites without calling for intensive computer resources. Several neural networks were applied and compared to interpret well logs, including layer picking, forward simulation, and inversion. A modular neural network was trained to extract the layer boundaries from a single unfocused tool response. The network showed the capability to pick layer boundaries but the confidence level was not uniformly high. Five different networks were trained with multiple tool responses to pick the layer boundaries. The results showed that the modular neural network and resilient back-propagation network could produce the most accurate results. The picked layer boundaries using multiple tool responses were located at a higher confidence level and less noise than using a single response tool. Fast forward modeling was performed with a modular neural network. The trained neural networks indicated that the modular neural network could predict the forward responses with an average accuracy of above 95%. An error analysis suggested that the neural network errors could be approximately described by a Gaussian distribution. A sensitivity test was also investigated to analyze how the errors would propagate back to the formation resistivity estimations. Larger errors were produced in conductive and thin layers, but smaller errors in thick layers. A pattern recognition-based fast forward modeling was developed. A self-organizing network was employed to classify the whole data population into different classes. Twenty percent of the training patterns from each prototype class were used for training. A modular neural network was investigated to invert Geonics EM39 induction logs. Four sub-networks were generated based on the pattern of resistivities in three-layer models. The well logging curves were subdivided into segments, which represented three-layer models, and each sub-network estimated the resistivity and thickness of every layer. The network was tested by synthetic and field data and the results were very encouraging.
Degree ProgramGraduate College
Mining and Geological Engineering