The model that decides which mineral you found

A mineral percentage in an announcement is usually a classification, not a measurement — and that is precisely where machine learning enters minerals work.

Reviewed by Peter Uppal

A circular microscope field filled with tightly packed mineral grains in cross-section, pale, dark and brown, each separated by thin boundaries. — Illustrative artwork: a polished mineral section of the kind examined in mineralogical analysis. Not a facility, equipment or material connected to this project. · Illustration · Osmond Hub

The short version

When a mineral percentage appears in an exploration announcement, it is usually not a direct measurement of that mineral: it is either a calculation from bulk chemistry or a classification made by matching an instrument's output against a reference library. Machine learning enters minerals work most consequentially at that step, and the published limits of the technique are limits of the library and the training data rather than of the algorithm. For an assemblage where titanium sits in more than one mineral and the rare earths sit in more than one host, the classification is not a formality — it is where the number comes from.

The number that is not a measurement

Start with a figure. Osmond reports Zone 1 bulk channel samples at 13.36% to 13.49% rutile and 1.56% to 1.62% monazite(opens in a new tab). Those are mineral percentages, and the company's own qualification travels with them: mineral percentages in its reporting are mass-balance estimates rather than direct measurement, a caveat stated in the appendix of its 16 July 2026 release. Nobody counted rutile grains. Something calculated how much rutile would account for the titanium the assay found.

That procedure has a name in the literature. A published account defines it directly: element-to-mineral conversion is the process of converting bulk chemistry data into mineral grades using known mineral chemistry(opens in a new tab), resting on the principle that the mineral grades multiplied by the chemistry of the minerals is equivalent to the bulk rock chemistry(opens in a new tab). It is a solved linear problem — provided you know which minerals are present and what each is made of.

The same source is candid about where that assumption fails. Elements can only be assigned to minerals once, making it difficult for such an approach to differentiate between minerals that share a similar chemistry and thus an overlap in the assigned element(opens in a new tab), and the mass balance approach is limited to deposits with simple mineral assemblages(opens in a new tab). That study concerns a lithium deposit and is cited here for the method rather than the mineralogy, but the constraint transfers. In a rutile-ilmenite-zircon-monazite assemblage, titanium has at least two homes and the rare earths have at least two — and one of those homes is not a fixed species: iron is leached from ilmenite during weathering, upgrading the TiO₂ of what remains and grading it towards leucoxene. No retrieved extract of the USGS account carries that, so it stands as general mineral-sands geology, not as a USGS finding. A calculation that must assign titanium to "rutile" or "ilmenite" is drawing a line across that series.

The other route, and its own failure mode

The alternative to calculating mineralogy is measuring it, with automated mineralogy on a scanning electron microscope. That is a classification problem, and it is worth seeing exactly how the classification is made.

A technical account of one such system describes it plainly: any unknown spectra collected from a point or a segment, together with related information from other detectors, are used to find the closest or first match to the records in the database providing mineral/phase identification(opens in a new tab). The identification is a nearest-match against a library. Everything depends on the library being right for the rock in front of it.

The same paper quantifies the ambiguity in the underlying signal: among the roughly six thousand known minerals, there are over 700 cases of overlapping BSE image data because of differences in average atomic numbers being below 0.1 Z(opens in a new tab), and the paper is explicit that BSE alone cannot reliably discriminate many phases with similar mean atomic numbers, which is why it holds that the use of combined BSE and EDS is strongly advisable(opens in a new tab). A separate study of phase-library practice puts the operational version of the problem: generic libraries often fail to capture site-specific mineral compositions, resulting in misclassification and unclassified pixels, particularly in systems with solid solution behaviour, compositional zoning, and textural complexity(opens in a new tab), and classification accuracy depends on the quality of the user-defined phase library(opens in a new tab). Even a species that is in the library can be missed: the composition of that species in an actual sample may lie outside the rules that were adapted to the library material, so the mineral in the sample may be misidentified or remain unidentified(opens in a new tab).

The interesting question is not whether a model can name a mineral. It is what happens to the number in the announcement when it names one wrong.

What machine learning actually contributes, at the size it has been shown

This is the point where machine learning has the clearest published case in minerals work, and also where the published results are narrower than the headline suggests.

One study trained a convolutional neural network to interpret hyperspectral core-scanning data against SEM-based automated mineralogy as ground truth. It reports: we obtained a prediction accuracy of 80%(opens in a new tab), and under ten-fold cross-validation an accuracy of 78%(opens in a new tab), rising so that as the votes of more pixels are combined an accuracy above 90% is eventually reached(opens in a new tab).

Those numbers should be quoted only with their scope attached. The model was trained on four mineral classes — quartz, K-feldspar, Na-K-feldspar and muscovite — on a 75/25 train-test split. That is a common-silicate problem, not a heavy-mineral assemblage, and the accuracy figure does not transfer to one. The authors attach two further cautions in their own words: this type of class imbalance can result in predictions that may provide misleading results, as a classifier can simply predict the majority class(opens in a new tab), and while our CNN model provided good mineral predictions we cannot easily say what spectral characteristics were useful to the network in reaching these identifications(opens in a new tab).

The second of those is the one that matters for reporting. A method whose reasoning cannot be inspected is a method whose errors cannot be anticipated, only measured after the fact against something else.

Where the judgement still has to sit

None of this makes classification unreliable; it makes it a stage with a stated uncertainty, like an assay. What it rules out is treating the output as an observation.

That is also what the disclosure system insists on. Osmond's exploration results are signed by a named Competent Person, with a second named Competent Person for mineral processing. A model can propose that a grain is rutile. A person still has to stand behind the sentence saying the deposit contains rutile, and to state — as Osmond does — the basis on which the percentage was arrived at. An inference layer makes more of the rock legible. The discipline is that legibility and evidence are not the same thing.

Exploration results and mineralogical estimates only. Orión has no JORC-compliant Mineral Resource or Reserve; maiden MRE and Scoping Study pending, targeted Q3 CY26.

Sources

Journal of Geosciences, 2018 — Automated mineralogy and petrology: applications of TIMA(opens in a new tab)
Minerals (MDPI) 15(11) 1118 — Automated, Not Autonomous: refining and validating phase libraries(opens in a new tab)
Minerals (MDPI) 16(2) 139 — Predictive Modelling of Lithium Mineral Grades from Chemical Assays(opens in a new tab)
Geosciences (MDPI) 13(7) 192 — Interpretation of Hyperspectral SWIR Core Scanning Data Using SEM-Based Automated Mineralogy(opens in a new tab)
USGS SIR 2010-5070-L — Deposit model for heavy-mineral sands in coastal environments(opens in a new tab)
Osmond figures: osm_grade — 19 February 2026; releases of 16 July 2026 and 14 August 2026 (ASX:OSM)

Related reading

What a thorium channel can and cannot see covers the instruments that generate much of the data this layer interprets.
Reconciling a mine that makes four products follows the same numbers to the point where they are checked against what a plant actually produced.
How two properties sort four minerals explains why distinguishing these particular species is not an academic exercise.
The monazite upgrade: what the preliminary testwork found (Science · Inside the testwork) shows a head grade and a concentrate grade being reported on two different bases.