Quantifying and modelling fractured subsurface rocks, characterised by their complex geometric heterogeneity, is crucial to the geo-energy transition because it helps predict flow properties in fractured systems. Multiscale Digital Rock Technology (MDRT) offers a solution to analyse comprehensive fluid flow mechanisms from the pore scale to much larger scales. In addition, artificial intelligence (AI) techniques can add significant value to geoscience workflows, automating time-consuming tasks, some even prohibitively long if done manually (such as 3D image volume labelling), and obtaining new insight from combining highly diverse data sources.
We propose a novel machine-learning algorithm for semantic segmentation of rock matrix, fractures, vugs, and secondary mineralogy. After implementing and examining deep and shallow-learning approaches, we concluded to use shallow machine-learning methods for increased computational efficiency and explainability while achieving comparable accuracy. By integrating our novel machine-learning algorithm into the multiscale Pore Network Model (PNM) code, we improve the modelling method of subsurface flow, particularly in complex fractured subsurface systems and carbonates.
The resulting algorithm accurately discriminates between pores, fractures, and vugs. Therefore, it enhances the accuracy of pore-fracture-vug network extraction and simulation and provides an improved analysis of complex rock structures. Moreover, the segmentation results are integrated into a Fracture-Pore Network Model, validated against high-fidelity OpenFOAM simulation.
This integration of fractures into the PNM code allows for larger scale fluid flow simulation in complex fractured subsurface systems. The current research produced a fast algorithm that accurately and automatically segments X-ray micro-computed tomography (micro-CT) samples having pores, fractures, and vugs. Our validation also showcases the potential of this algorithm to improve existing industrial core analysis practices.