Scale (ratio)

Geotechnical engineering

Materials science

Cartography

Binary image

Image processing

Computed tomography

Grayscale

Computer science

Surface (topology)

Geology

Radiology

Porosity

Physics

Pattern recognition (psychology)

Mathematics

Optics

Geography

Image (mathematics)

Artificial intelligence

Geometry

Porous medium

Composite material

Medicine

X-ray

Artificial neural network

Binary number

Convolutional neural network

Arithmetic

Journal of Petroleum Science and Engineering

In this paper, Convolutional Neural Networks (CNNs) are trained to rapidly estimate several physical properties of porous media using micro-computed tomography (micro-CT) X-ray images as input data. The tomograms of three different sandstone types are subdivided to create a dataset consisting of 5,262 training images and 2,000 testing images. Porosity, specific surface area, and average pore size of each image are computed. The proposed CNN framework is trained with binary images and greyscale images separately and the corresponding computed properties. The results from testing the model are promising as the relative error in determination of porosity, surface area and average pore size is less than 6% when the model is trained with binary images and less than 7% when greyscale images are used. Other aspects related to model training and optimisation are discussed.

Machine learning for predicting properties of porous media from 2d X-ray images