Many unconventional reservoirs exhibit a high level of vertical heterogeneity in terms of petrophysical and geo-mechanical properties. These properties often change on the scale of centimeters across rock types or bedding, and thus cannot be accurately measured by low-resolution petrophysical logs. Nonetheless, the distribution of these properties within a flow unit can significantly impact targeting, stimulation and production. In unconventional resource plays such as the Austin Chalk and Eagle Ford shale in south Texas, ash layers are the primary source of vertical heterogeneity throughout the reservoir. The ash layers tend to vary considerably in distribution, thickness and composition, but generally have the potential to significantly impact the economic recovery of hydrocarbons by closure of hydraulic fracture conduits via viscous creep and pinch-off. The identification and characterization of ash layers can be a time-consuming process that leads to wide variations in the interpretations that are made with regard to their presence and potential impact. We seek to use machine learning (ML) techniques to facilitate rapid and more consistent identification of ash layers and other pertinent geologic lithofacies. This paper involves high-resolution laboratory measurements of geophysical properties over whole core and analysis of such data using machine-learning techniques to build novel high-resolution facies models that can be used to make statistically meaningful predictions of facies characteristics in proximally remote wells where core or other physical is not available. Multiple core wells in the Austin Chalk/Eagle Ford shale play in Dimmitt County, Texas, USA were evaluated. Drill core was scanned at high sample rates (1 mm to 1 inch) using specialized equipment to acquire continuous high resolution petrophysical logs and the general modeling workflow involved pre-processing of high frequency sample rate data and classification training using feature selection and hyperparameter estimation. Evaluation of the resulting training classifiers using Receiver Operating Characteristics (ROC) determined that the blind test ROC result for ash layers was lower than those of the better constrained carbonate and high organic mudstone/wackestone data sets. From this it can be concluded that additional consideration must be given to the set of variables that govern the petrophysical and mechanical properties of ash layers prior to developing it as a classifier. Variability among ash layers is controlled by geologic factors that essentially change their compositional makeup, and consequently, their fundamental rock properties. As such, some proportion of them are likely to be misidentified as high clay mudstone/wackestone classifiers. Further refinement of such ash layer compositional variables is expected to improve ROC results for ash layers significantly.
