Microanalysis techniques like Scanning Electron Microscopy (SEM) and Quantitative Electron Microscopy (QEM) offer valuable insights into material properties. However, data acquisition can be time-consuming and expensive, and measurements are constrained to a flat surface. This study explores the potential of machine learning to predict microanalysis images and microanalysis-derived measurements (such as porosity) in 3D directly from Computed Tomography (CT) data.

Classical image segmentation approaches cannot differentiate between materials in a CT image that are not separable based on their intensity alone. Whereas machine learning-based methods can learn to differentiate such materials based on textures and other contextual information. This study demonstrates that machine learning can be effective for the characterisation of materials such as iron ore, which are difficult to segment using conventional approaches.