Integrated workflow for prediction of organic pore volume in unconventional plays, an example from the Duvernay formation, Canada

Abstract

Organic pores provide the primary storage space for hydrocarbons in some unconventional plays. However, organic pore volume and pore size distribution data are not routinely collected due to time, labor, and cost. This work presents an efficient workflow for the estimation of organic pore volume in self-sourcing reservoirs using more routinely gathered mineral and geochemical data and machine learning methods. This approach provides comparable results to the analytical approach of using subcritical N₂ adsorption, but at significantly reduced cost. The Late Devonian Duvernay Formation of western Canada is used as an example to develop the workflow. This workflow should be adaptable to other locations.

This work utilized total organic carbon (TOC), Rock-Eval pyrolysis, and mineral data. Data processing was performed prior to modeling to improve prediction accuracy and precision. Specifically, data transformation, stratification, and stratified three-fold cross validation approaches are used to overcome limitations of small datasets and improve model optimization. Multilinear Regression and Random Forest modeling are benchmarked for prediction optimization. Ensuring that training datasets include end-member data is critical to increase the reliability of model generalization. Stepwise regression and factor significance are used to select important factors in the modeling, observing that not all available data are needed for a meaningful prediction.