Deep Learning for a Fast and Accurate Prediction of Complex Carbonate Rock Permeability From 3D Micro-CT Images

Abstract

Abstract Correctly predicting subsurface flow properties is critical in many applications, ranging from water resource management to the petroleum industry. In the present paper, we establish a workflow to apply machine and deep learning (DL) to quickly and accurately compute petrophysical properties based on micro-CT images without any computationally intensive procedures. The pore network modeling (PNM) approach is widely used for fast computation of flow properties, albeit with less accuracy due to the inherent simplification of the pore space. Alternatively, direct simulation, such as the lattice Boltzmann method (LBM) is very accurate; however, its high computational cost prevents this approach from including all the relevant flow physics in a single simulation. After assessing numerical techniques ranging from PNM to the LBM, a framework based on machine learning (ML) is established for a fast and accurate prediction of permeability directly from 3D micro-CT images of complex Middle-East carbonate rock. We use thousands of samples from which engineered features—based on pore network modelling and images analysis—are fed into both shallow and deep learning algorithms to compute, as an output, the permeability in an end-to-end regression scheme. Within a supervised learning framework, algorithms based on linear regression, gradient boosting, support vector regression, and convolutional neural networks are applied to predict porous rock petrophysical properties from 3D micro-CT images. In addition, a hybrid neural network accounting for both the physical properties and 3D raw images is investigated. Finally, the estimated permeability of a complex carbonate by ML is found to be in good agreement with a more intensive simulation by voxel-based direct simulation. Furthermore, a significant gain in computational time—approximately three orders of magnitude—is achieved by applying ML compared to the LBM. This work highlights the critical role played by features engineering in predicting petrophysical properties using DL. The proposed workflow, combining deep learning and rock imaging and modeling, has great potential in reservoir simulation and characterization to swiftly and accurately predict petrophysical properties of porous media.