Quantification of the microstructural properties of CCS and radioactive waste target rocks using Convolutional Neural Networks

Digital rock models are becoming increasingly important in addressing the challenges of transitioning to sustainable energy. While traditionally employed to model fundamental petrophysical and geomechanical processes, their utility is expanding into critical applications such as carbon capture and storage (CCS), geothermal energy development, and subsurface energy storage. By using advanced imaging, simulation, and multi-scale analysis techniques, digital rock models provide a detailed understanding of pore-scale properties and their implications for fluid flow, geomechanics, and geochemistry. These insights are essential for optimizing low-carbon energy systems and ensuring reservoir integrity during energy storage and CO₂ sequestration. This work highlights some of the recent advancements in digital rock technologies and their contributions to innovative solutions in sustainable energy development.

Estimating the physical properties of rocks, a crucial and time-consuming process in the characterization of geothermal reservoirs, CCUS, and other renewable energy resources, has seen a shift from traditional laboratory experiments to the increasing use of digital rock physics. A key requirement of many forms of pore structure image analysis is that they require binary images to distinguish pore-space from non-pore-space (mineral phases). These are often obtained by thresholding grayscale SEM or X-ray tomographic images. In this study, we present the collection and processing of exceptionally high-quality two-dimensional images of carbonate rocks, with a resolution of 16-bit density and dimensions of 29056 × 22952 pixels. This dataset, subdivided into 155 smaller images of 2048 × 2048 pixels each, was further enhanced using data augmentation techniques such as rotation and reflection, creating a diverse and non-redundant set of training images.

The objective of this work is to train a machine-learning model capable of predicting porosity directly from the images. A convolutional neural network (CNN) was developed and modified for this purpose, using 60% of the dataset for training. The training process involves pre-labeled images, which are used to optimize the weights of the neural network. So far, the CNN has achieved an accuracy of 89.55% in predicting porosity during the training phase. Validation and testing datasets were employed to evaluate and refine the model’s performance, with ongoing efforts aimed at surpassing 95% accuracy in testing. Furthermore, we are working on analyzing the relational characteristics of porosity to expand the applicability of this approach. Initial work in 2D and 3D that has the power to discriminate between mineral phase, between connected and unconnected porosity, and to quantify the pore fluid-mineral surface area, are also in progress. This latter property is extremely relevant to CCS targets where the area for CO2 adsorption is an important parameter which is difficult to assess.

This research not only enhances our ability to quantify key petrophysical properties but also contributes to the development of sustainable energy technologies. The work has significant potential to enhance geothermal resource evaluation and advancing carbon capture and storage (CCS) initiatives, playing a critical role in the transition to low-carbon energy solutions.