Machine Learning for Rapid Lithotype Classification from Multi-Log Suites to Assist Interpretation and Property Modeling
Multiple petrophysical well logs are often acquired in well bores drilled for oil and gas development. A very common use of these suites of logs is well-to-well correlation of geologically significant events to map structural and stratigraphic spatial variations, especially in vertically and laterally complex sequences. Additionally, multivariate analysis of multiple logs may be used to predict other log properties not originally recorded in the well. Combinations of log properties can be diagnostic of the lithology and fluid-content of rock types in the zone of interest. Determination of major lithotypes from log suites can greatly facilitate interpretation and lithotype-dependent property estimation. This determination is especially applicable in unconventional resource plays of highly variable lithostratigraphy in which there are tens of thousands of well bores, where the number and type of logs can vary from well to well. However, generating lithotype logs from so much and such highly variable data has often been a daunting and time-intensive task, especially given the short timeframe of many business opportunities. This paper presents a workflow of unsupervised (e.g., Kmeans), semi-supervised (e.g., hierarchical), and supervised (e.g., discriminant analysis and decision trees) classification tools for rapidly modeling lithotypes and their associated likelihoods. Considerations of the number and type of logs that vary across the study area and of the number of classes and relationships among the classes are explored and illustrated. Modeling begins at well bores with the greatest diversity of log types and then proceeds to other wells with fewer and fewer logs. Log aliases prove a powerful tool for prioritizing the results with different degrees of model confidence. Finally, the resulting classifications are used to build multivariate models (e.g., Alternating Conditional Expectation and neural networks) for predicting rock properties in which these models are class dependent. Examples from the Niobrara-Codell sequence in the Denver-Julesburg Basin and the Bone Spring - Wolfcamp sequence in the Delaware Basin are given.
AAPG Datapages/Search and Discovery Article #90350 © 2019 AAPG Annual Convention and Exhibition, San Antonio, Texas, May 19-22, 2019