Using machine learning to discriminate between mineral phases and pore morphologies in carbonate systems

Digital rock models are becoming an essential tool not only for the modelling of fundamental petrophysical processes, but in specific key applications, such as Carbon Capture and Underground Storage (CCUS), geothermal energy exploration, and radioactive waste storage. By utilizing advanced imaging and simulation techniques, digital rocks provide indispensable insights into the porous structures of geological formations, crucial for optimizing CO2 storage, enhancing geothermal reservoir characterization, and ensuring the secure containment of radioactive waste. This abstract aims to present new advances using digital rocks to study these pressing environmental and energy challenges.Estimating the physical properties of rocks, a crucial and time-consuming process in both the characterisation of hydrocarbon, geothermal and CCUS resources, has seen a shift from traditional laboratory experiments to the increasingly prevalent use of digital rock physics. A key requirement of many forms of pore structure image analysis is that they require binary images showing pore-space vs. non-pore space (mineral phases). These are typically obtained by thresholding grey scale SEM or X-ray tomographic images to separate the two phases. In this paper, we have adapted a 2D process-driven MATLAB model to generate synthetic porous media images, laying the foundation for simulating authentic SEM images. The objective of the computational framework outlined in this study is to train a machine-learning model capable of predicting various types of porosity. Drawing inspiration from recent advances in machine learning applied to porous media research, our approach involves the development of deep learning models utilizing Convolutional Neural Networks (CNN). Specifically, we aim to quantitatively characterize the inner structure of the 2D porous media based on their binary images through the implementation of these CNN models. This framework consists of: (i) Generating synthetic porous media images through a process-driven model, (ii) training a neural network that takes a labelled synthetic image as input and gives two types of porosity as output, (iii) whereupon the trained model can be applied to provide types of porosities for new images that are not in the training database. The generated data are divided into training, validation, and testing datasets. The training dataset optimizes CNN parameters for accuracy, the validation dataset aids in hyperparameter selection and prevents overfitting, and the testing dataset evaluates the predictive performance of the trained CNN model.This research not only advances the understanding of fundamental geological processes but also plays a crucial role in optimizing the utilization of renewable energy sources such as geothermal and contributing to the effective management of carbon capture and storage initiatives.