Advanced Microscopy and Machine Learning for Multiscale Analysis of Porosity and Mineralogy

Recent advances in microanalytical imaging and machine learning enable quantitative, multiscale characterization of geological materials with direct relevance for subsurface energy storage. This study presents an integrated workflow combining Broad Ion Beam (BIB) sample preparation, Scanning Electron Microscopy (SEM), Energy Dispersive X-ray Spectroscopy (EDX), and advanced machine learning to quantify pore structures, mineralogy, and their spatial relationships from the micrometre to nanometre scale (Klaver et al. 2021).

High-resolution secondary electron (SE2) and backscattered electron (BSE) imaging, complemented by low-resolution EDX data, provides multimodal datasets for automated analysis. Pore networks are segmented using a pre-trained U-Net deep learning model, enabling efficient and accurate porosity quantification. Mineralogical phases are identified and quantified through a semi-automatic, decision-tree–based segmentation approach. The alignment of SE2 and BSE datasets allows porosity to be directly correlated with specific mineral phases, establishing a robust link between microstructure, mineral composition, and petrophysical properties (Jiang et al, 2021).

The applicability of this technology-driven approach is demonstrated through two case studies. Case study 1 investigates geological hydrogen storage in underground salt caverns, focusing on the impact of biotic and abiotic reactions on anhydrite. Flow-cell experiments combined with cryogenic BIB-SEM analyses enable early detection of microstructural, mineralogical, and pore-space changes induced by hydrogen, hydrogen sulfide, and microbial sulfate reduction. Despite slow reaction kinetics, microstructural observations reveal the substantial onset of chemical alteration, biofilm formation, and evolving pore connectivity at the submicron scale, providing essential constraints for geochemical and hydraulic models (Berest et al., 2024).

Case study 2 examines fault sealing in mechanically layered limestone–marl successions. Oriented transfer samples from normal fault systems were analysed using multiscale microanalytical workflows to capture marl smearing, mechanical mixing, fracturing, and cementation processes. High-quality microstructural datasets serve as ground truth for training machine learning algorithms for efficient interpretation of 2D image data. The results show that fault cores are composed of recurrent structural building blocks whose distribution and sealing capacity are strongly controlled by the presence and properties of marly interbeds (Schmatz et al., 2022).

Overall, the integrated microscopy–machine learning framework provides a transferable, data-driven approach for quantifying coupled structural, hydraulic, and geochemical processes in complex geological systems.

References

Berest et al.,2024. Risk assessment of hydrogen storage in a conglomerate of salt caverns in the Netherlands. KEM-28 report. https://www.kemprogramma.nl/documenten/2024/04/03/kem-28-project-rapportfinal-report-kem-28-h2c3-240403_v2

Jiang et al., 2021.Workflow for high-resolution phase segmentation of cement clinker from combined BSE image and EDX spectral data. Journal of Microscopy, 1-7.

Klaver et al., 2021. Automated carbonate reservoir pore and fracture classification by multiscale imaging and deep learning. 82nd EAGE Annual Conference & Exhibition, Oct 2021, Volume 2021, p.1 – 5.

Schmatz et al., 2022. Prediction of Fault Rock Permeability With Deep Learning: Training Data from Transfer Samples of Fault Cores. 83rd EAGE Annual Conference & Exhibition, Jun 2022, Volume 2022, p.1 – 5.