Eventos Anais de eventos
ENCIT 2022
19th Brazilian Congress of Thermal Sciences and Engineering
Permeability prediction of karst porous media using Convolutional Neural Networks
Submission Author:
Sergio Ribeiro , RJ
Co-Authors:
Gabriel Grenier Ferreira Motta, Monique Feitosa Dali, Sergio Ribeiro, Marcio CARVALHO
Presenter: Sergio Ribeiro
doi://10.26678/ABCM.ENCIT2022.CIT22-0229
Abstract
Averaging methods are widely used to estimate the average permeability of oil reservoirs. The Darcy Equation receives a measured flow, corresponding to an applied pressure differential, to obtain a good approximation of mean permeability. For media with homogeneous porous structure, it is possible to obtain a relationship between permeability and porosity of the media. However, karstified rock formations, such as carbonates, present vugs, fractures and other cavernous structures with abrupt variations in local porosity and permeability. Determining a good correlation between the patterns of karst structures and the generated increase in permeability would considerably increase the accuracy of heterogeneous reservoir characterization methods. Recently, convolutional neural networks (CNNs) have been used to estimate the effects of karst characteristics on the equivalent permeability of a porous carbonate medium, based on images of the macroporosity structures of the sample (Dali et al., 2020). The methodology proposed by Dali et al. (2020) consists of: (i) simulating the flow through a 2D mesh built on the basis of rock microtomography images using the Brinkman model; (ii) creation of a training database from the macroporosity binary images and the corresponding increase in simulated permeability in relation to the permeability of the rock matrix; (iii) finally the trained CNN is used to estimate the equivalent permeability increment of unseen porous karst media images. The present work re-implements this methodology in two-dimensional microtomography images obtained from Brazilian carbonate samples. In addition, it presents a comprehensive sensitivity analysis of the CNN architecture and training parameters with respect to the effective accuracy of the permeability prediction. Finally, the current study complements the previous study with an assessment of CNN’s ability to generalize patterns learned from an image-specific training dataset. The analysis evaluates which network characteristics allow to reasonably predict the permeability increments within a test dataset composed of unseen rock samples with different permeability ranges from the training ones.
Keywords
machine learning, Equivalent permeability, Reservoir Characterization, Karst porous media
