A Novel Method of Automatic Training Data Selection for Estimating Missing Well Log Zone Using Neural Networks
Yu, Yingwei¹; Seyler, Douglas²; McCormack, Michael D.³
¹IHS, Houston, TX.
²Blueback Reservoir Americas, Houston, TX.
³Olympic Geoscience Consulting, Sequim, WA.
Well log interpolation is of great importance to understand the lithology and history of a sedimentary basin. A neural network-based algorithm is presented that estimates missing intervals or corrupted zones of log curves using log curves from the same borehole and other nearby wells. In general, for approaches using neural networks, the accuracy of the prediction is highly dependent on the quality of the training data. The user has to select the training data manually, but this method is time consuming, and sometime can be inaccurate. A unique feature of this algorithm is the extensive analysis and preprocessing of all candidate wells and their log curves to determine the set of wells, curves, as well as log data samples within each curve, that will yield the best estimate of the missing log interval. This refined set of curves and data samples are used to train a neural network which then estimates the missing log interval. An additional output of this analysis is a confidence value for each estimated log sample which provides a qualitative measure of the accuracy of the neural network prediction.
AAPG Search and Discovery Article #90155©2012 AAPG International Conference & Exhibition, Singapore, 16-19 September 2012