Probabilistic three-dimensional mineralogical characterisation of drill core using μXRF and XCT: application to the Nautanen North IOCG deposit
This study presents a probabilistic, deep learning-based framework that integrates 2D micro-XRF and 3D X-ray computed tomography data to generate accurate, uncertainty-aware three-dimensional mineralogical maps of drill cores, thereby enhancing geometallurgical characterization of the Nautanen North IOCG deposit.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Deep underground, where the earth's crust holds the promise of copper, gold, and other metals, geologists rely on long, cylindrical samples of rock called drill cores to understand what lies beneath the surface. To make sense of these rocks, they need to know not just what elements are present, but exactly which minerals hold them, how those minerals are arranged, and how they interact with one another. This detailed knowledge, known as geometallurgy, is crucial because the way a rock breaks apart and releases its valuable metals depends entirely on its internal texture and composition. Traditionally, scientists have had to choose between two imperfect methods: chemical tests that tell them the total amount of an element but destroy the rock's structure, or high-resolution scans that show the surface in detail but cannot see inside the rock without cutting it apart. The challenge has always been to connect the detailed surface view with the hidden three-dimensional interior without destroying the sample.
A team of researchers has developed a new way to solve this puzzle by combining two powerful imaging techniques with a smart computer model. They took four samples of drill core from the Nautanen North deposit in northern Sweden, a site known for its complex mix of minerals and intense geological history. First, they used a specialized scanner called micro-X-ray fluorescence to map the chemical composition of the rock's exposed surface, creating a detailed two-dimensional map of the minerals present. Then, they used X-ray computed tomography, a technology similar to a medical CT scan, to create a full three-dimensional digital model of the same rock samples, showing how X-rays pass through the interior. The breakthrough came when they taught a computer program to recognize the patterns in the three-dimensional scan that corresponded to the mineral maps on the surface. By learning from the surface, the computer learned to predict what minerals were hiding just beneath the skin of the rock, effectively extending the two-dimensional map into the third dimension.
The researchers found that this method works well for some minerals but struggles with others, depending on how distinct their internal signatures are. The computer model became highly accurate at identifying the most common minerals, which form the bulk of the rock, and could also reliably spot large, distinct crystals of garnet and quartz. However, it faced significant difficulty distinguishing between different types of iron-bearing minerals, such as various sulfides and oxides. This is because these minerals often look very similar to the X-ray scanner, producing overlapping signals that are hard to separate, especially when they are mixed together in fine grains. Instead of forcing a single, potentially wrong answer, the system was designed to provide a probability, telling the geologist how confident it is in its prediction and where the uncertainty lies. This approach allows scientists to see not just a static map, but a dynamic view of where the rock's composition is clear and where it remains ambiguous.
The study demonstrated that this probabilistic framework could successfully predict mineral distributions throughout the volume of a drill core, not just on the surface. When the researchers applied the trained model to the interior of a rock sample, the predictions remained consistent and coherent as they moved deeper into the rock, maintaining the spatial patterns observed on the surface. The system generated a three-dimensional volume filled with probability fields, showing which mineral groups were likely present at every point inside the core. While the model was most reliable in the stable interior of the rock, its confidence naturally decreased near the curved back edge where the rock sample ended, a behavior that correctly reflected the limitations of the data rather than a failure of the method. This ability to generate a continuous, three-dimensional view of mineralogy from limited surface data offers a powerful new tool for mining companies. It allows them to identify valuable zones and potential processing challenges deep within a core without having to cut the rock apart, preserving the sample for further testing while providing a much richer understanding of the ore body's internal structure.
The success of this method relies on the fact that different minerals interact with X-rays in unique ways, creating a signature that the computer can learn to recognize. However, the researchers were careful to note that this is not a perfect identification system for every single mineral. The model groups minerals into broader categories based on how they behave in the scan, rather than trying to name every specific species, which would be impossible given the overlapping signals. For instance, the system treats a mix of iron oxides and sulfides as a single group because their X-ray responses are too similar to separate reliably at the scale of the scan. This limitation is not a flaw in the technology but a realistic acknowledgment of the physics involved. By focusing on these broader, more distinguishable groups, the researchers created a system that is robust enough to work across different rock samples, even when the specific mix of minerals changes from one hole to another.
The study also highlighted the importance of how the data is prepared and validated. The researchers used a rigorous testing method where they trained the computer on three rock samples and tested it on a fourth one it had never seen before. This ensured that the model was learning general rules about the rock's structure rather than just memorizing the specific patterns of the training samples. The results showed that the model could indeed transfer its knowledge to new, unseen rocks, correctly identifying the dominant mineral groups in the test samples. This cross-sample reliability is essential for practical application, as geologists need to trust that the patterns learned from one drill hole will hold true for the next. The study confirmed that while the computer can learn to see the hidden structure of the rock, it must be guided by real, physical data from the surface to do so accurately.
Ultimately, this work provides a bridge between the detailed, two-dimensional world of surface mineralogy and the complex, three-dimensional reality of the underground ore body. It does not replace the need for detailed chemical analysis or physical testing, but it offers a way to fill in the gaps between those tests. By generating a continuous, three-dimensional map of mineral groups along with a measure of confidence for each prediction, the framework gives geologists a new way to visualize the variability of the rock. This helps in defining geological zones more accurately and selecting the most representative samples for further testing. The ability to see inside the rock without destroying it, while understanding the limits of what can be seen, represents a significant step forward in how we characterize the resources hidden beneath our feet.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.