Rapid Rock Properties Estimation from Micro-CT Images by Deep Leaning

Abstract

Abstract Digital rock physics (DRP) can compute several rock properties such as porosity, tortuosity, formation factor, permeability, thermal conductivity, and elasticity. To calculate these properties, the DRP workflow usually includes micro-CT imaging, processing the images (filtration and segmentation), and simulation of physical processes. These steps can be time-consuming and prone to human error. In this study, three image datasets of a sandstone sample were used. The first dataset includes the original micro-CT images without any processing. The second dataset contains micro-CT images filtered by the non-local means filter to reduce the noise. The third dataset was created by segmenting the filtered micro-CT images. From these segmented images, porosity and tortuosity were calculated. Tortuosity was computed using pore network modeling (PNM). Three deep convolutional neural networks (DCNNs) were developed to predict tortuosity and porosity from the image datasets. The accuracy of the predictions was assessed by the average absolute percentage error (AAPE) and correlation coefficient (R). The predictions from the three image datasets are comparable. R and AAPE of the predicted porosity by DCNNs are about 0.99 and (1%-1.8%). R and AAPE of the predicted tortuosity by DCNNs are about 0.93 and (5.9%-6.2%). The results show that simple rock properties (porosity) are more accurately predicted by DCNN than advanced rock properties (tortuosity). The predictions from the micro-CT images (without any processing) are almost as accurate as the predictions from the filtered and segmented images. This could be due to the use of deeper DCNN. Therefore, DCNN can quickly predict some rock properties from unprocessed micro-CT images without significant loss of accuracy. This paper shows that DCNN has the potential to eliminate the need for image processing and simulation in the DRP workflow, which may allow semi-automation of the DRP workflow. Additionally, micro-CT scanner manufacturers may be able to provide end-to-end solutions. In this case, raw data (unreconstructed projections) could be used with micro-CT images to enhance the accuracy of the predicted rock properties.