Predicting carbonate formation permeability using machine learning

Abstract

It is imperative to characterize the formation permeability to simulate the flow behavior at subsurface conditions. An accurate characterization at the core scale is possible when large samples are available, but often this is not the case, as such samples are hard to recover. Instead, drill cuttings (small pieces) are usually the only source available, especially in real-time conditions. Thus, mercury injection capillary pressure measurements, which are applicable to small pieces, have been used to infer the formation permeability. The challenge is that capillary pressure measurements entail further interpretations, as they can be converted to the pore-throat size distribution but not directly to the permeability. Thus, researchers have proposed different empirical and theoretical relations to predict the permeability. The present study uses machine learning, a data-driven approach, to predict carbonate formation permeability. The data-driven approach does not impose any restriction on the spatial distribution of the pore-throat sizes in the network of connected pores, but rather trains models based on the existing data. The present study is based on 193 carbonate samples whose data (porosity, permeability, and mercury injection capillary pressure measurements) are available in the literature. The permeability values vary from nanodarcies to darcies. We propose two new correlations, with and without grouping analysis, for permeability prediction. The results are promising, as the averaged R2 score obtained with 50 iterations is larger than 0.96. The study provides a valuable tool for permeability prediction based on numerical methods that distinguish the pore structure by taking into account underlying trends in the measurements.