The study of geomaterial microstructures is fundamental to advancing our understanding of subsurface properties that affect fluid transport, reactivity, and mechanical integrity in geological systems. This thesis combines cutting-edge deep learning techniques with statistical tools from computational material science to characterise and reconstruct complex rock microstructures. Microstructures, specifically porosity, significantly influence the mechanical and transport properties of rocks, impacting processes ranging from geothermal energy extraction and carbon sequestration to fluid transport and mineral replacement. These processes play essential roles in diverse geoscientific and engineering applications, including modelling rock behaviour in the Earth, assessing geological hazards, and fluid-rock interactions. The primary goal of this work is to integrate deep learning, particularly generative adversarial networks (GANs), with interpretable statistical descriptors to achieve a detailed, quantifiable understanding of rock microstructures and their evolution, thereby providing enhanced tools for digital rock physics and addressing inherent limitations of common imaging techniques, specifically the trade-off between resolution and field of view. This thesis investigates the reconstruction and characterisation of 2D and 3D rock microstructures using GANs combined with statistical microstructure descriptors (SMDs). Advanced GAN models are employed to reconstruct complex 2D hydrothermal rock microstructures, evaluated using multi-point spatial correlation functions. This approach highlights the advantages of interpretable statistical measures over common deep learning metrics, enabling better assessment of model performance and detection of artefacts in the reconstructed microstructures. The work is extended to address the limitations of imaging techniques, particularly the trade-off between resolution and field of view. A comprehensive workflow based on a GAN with a specific architecture is proposed to reconstruct realistic 3D microstructures from 2D SEM and optical images. The results are validated against ground-truth micro-CT data and previous studies on sandstone samples used as benchmarks. The method proves valuable where features are too fine for conventional X-ray imaging or where sample heterogeneity cannot be captured by limited fields of view, emphasising the importance of generating diverse reconstructions for reliable property assessment. Building on these methods, the dynamic evolution of reaction-induced porosity is investigated using time-resolved (4D) synchrotron tomography of the KBr-KCl replacement system. Combining deep learning for image processing and SMDs, the study identifies a three-stage porosity evolution linked to transitions between advection- and diffusion-driven transport. These findings provide new insights into the interplay between pore structure evolution, dissolution-precipitation dynamics, and transport properties during fluid-rock interaction processes. Finally, the Pore-Edit GAN is developed to modify pore connectivity realistically, inspired by semantic image editing from computer vision. Trained on 4D experimental data of KBr-KCl replacement, the model is applied to reconnect isolated pores in a naturally altered meta-igneous rock where direct experimental observation is currently unfeasible. The generated reconstructions yield permeability estimates consistent with previous studies, suggesting that local permeability can increase significantly during mineral replacement. This work demonstrates the potential of generative models to modify and explore the influence of microstructural changes on rock properties, offering new pathways for predictive modelling and uncertainty quantification in geological systems.