Neural Network Modeling with Sparse Datasets

Abstract

Neural networks are important tools for the analysis and modeling of many types of petroleum data. Small datasets limit their utility, however, because of the need to provide separate training and testing datasets. We train neural networks so that all the data are used for training, model development, and testing. We use the training procedure on a network to analyze fracture spacing in the Lisburne Formation, northern Alaska. Analyzing the effect of bed thickness on the spacing, we find only a weak influence, with closer fractures in thick beds. This result agrees with statistical analysis of the Lisburne data, but is contrary to relationships reported elsewhere.

Keywords:

• activation function
• fracture spacing
• Lisburne

ACKNOWLEDGMENTS

Collection and analysis of the fracture data were funded by a grant by the U.S. DOE to the University of Alaska (Contract DE-AC26-98BC15102), T. Bui received additional support from Texas A&M University.