Rock Properties From Well Logs Using Implicit Predictive Modeling Technique to Reduce Massive Coring Operations in Oil-Sand Mining

Abstract

Oil-sand mining requires detailed characterization of the subsurface for safe and economic operations. Historically, this has been achieved via coring at a 100-200 m well spacing. The objective of this work is to reduce coring cost via better integration of wireline log data. Explicit predictive modeling techniques such as the Archie model are data-driven regression models based on core data where the equations and variables are explicitly expressed to describe the regression surfaces between predictors (well logs) and outcomes (rock properties). Unfortunately, explicit transforms between logs and rock properties are unavailable for some rock properties critical to oil-sand mining such as particle-size distribution (PSD) and cation exchange capacity from methylene blue index (MBI). Moreover, the generic regression surface of an explicit model can be too rigid to fit the core data within acceptable prediction errors. Implicit predictive modeling techniques use core and log data points (training dataset) to describe the regression surfaces between predictors and outcomes without explicit transforms. An implicit model can be built between any predictors and outcomes if a relationship is observed. This approach requires a sufficient core data to build an appropriate training dataset, which is ideally suited to oil-sand mining where core data are abundant. An implicit model has been built using gamma-ray, density and resistivity logs to predict critical rock properties: bitumen mass fraction, water saturation, PSD, and MBI. The K-nearest-neighbor method is used for its simplicity and performance. The predictors (logs), wells, and intervals for the training dataset are carefully selected to reduce noise and cover the full ranges of log and rock-properties. A rigorous workflow was developed and applied to examine model applicability, evaluate results, investigate sources of mismatches, and iteratively improve the model. This technique has been blind-tested against core data. Results suggest that many key attributes important to oil-sand mining can be predicted from wireline logs to an accuracy comparable to core data at the scale of the mine face (∼15 m). These results provided a technical basis for Kearl oil-sand mining operations to reduce the number of cores taken and save costs associated with field operations and laboratory analysis. The true measure of success can only be verified when mining operations excavate the area where the predictions were made.

AAPG Datapages/Search and Discovery Article #90259 ©2016 AAPG Annual Convention and Exhibition, Calgary, Alberta, Canada, June 19-22, 2016