Permeability Prediction using multivariant structural regression

Carl Zeiss Microscopy, 4385 Hopyard Rd., Pleasanton, CA, 94588, USA

^* Corresponding author: matthew.andrew@zeiss.com

Abstract

A novel method for permeability prediction is presented using multivariant structural regression. A machine learning based model is trained using a large number (2,190, extrapolated to 219,000) of synthetic datasets constructed using a variety of object-based techniques. Permeability, calculated on each of these networks using traditional digital rock approaches, was used as a target function for a multivariant description of the pore network structure, created from the statistics of a discrete description of grains, pores and throats, generated through image analysis. A regression model was created using an Extra-Trees method with an error of <4% on the target set. This model was then validated using a composite series of data created both from proprietary datasets of carbonate and sandstone samples and open source data available from the Digital Rocks Portal (www.digitalrocksporta.org) with a Root Mean Square Fractional Error of <25%. Such an approach has wide applicability to problems of heterogeneity and scale in pore scale analysis of porous media, particularly as it has the potential of being applicable on 2D as well as 3D data.