Permeability estimation from rock digital images segmented through deep learning

Abstract

Listen

Translate article

An accurate estimation of permeability is very important for the development and management of petroleum reservoirs. Knowledge of petrophysical properties is essential to achieve profitability in production in the oil and gas (O&G) industry. Modeling of flow properties at pore scales has gained great prominence in oil industry projects with the advances achieved with digital transformation. From digital images of reservoir rocks, it is possible to extract information from the porous system, in order to obtain petrophysical properties such as porosity, permeability, and, make suitable the comprehension of the connectivity of this system. The feasibility of acquiring three-dimensional images of these rocks, the industry began to use machine learning tools to assist the rock characterizations, thus, the entire modeling process is more efficient. And with high-performance computers, both the optimization and expansion of porous network have led to increasingly real flow calculations, which contributes to the knowledge of these properties. The objective of this research is to estimate the absolute permeability of rocks using micro tomography (microCT) images segmented using deep learning. It was used 12 samples of coquinas from the Morro do Chaves formation (Sergipe-Alagoas Basin), these samples are similar to those from the Itapema formation (Santos Basin). The images, with 42-micron resolution and in gray scale. Data were separated into training, validation and testing. Each batch of training images has 32 sub-images with 64x64 pixels (‘patch’) and the stride ratio equal to 1. It was used 100 learning epochs with an early stopping criterion. The model used for segmentation will be U-net convolutional neural network with cross entropy as cost function and Adadelta ([1]) optimization algorithm. Additionally, sensitivities were made regarding the use or not of the data augmentation technique and also regarding the use of input images with and without denoising filters. Deep learning models were generated in the Dragonfly program ([2]), and the modeling of the pore network, using the PNM algorithm (Pore Network Modeling).[1] Zeiler, Matthew D.. “Adadelta: an adaptive learning rate method” [Online] Available: https://doi.org/10.48550/arXiv.1212.5701[2] Dragonfly: commercial software [online] Available: https://dragonfly.comet.tech/