When dirt looks the same data decides how geochemistry and machine learning are revealing hidden rare earth riches beneath Western Australia’s surface

If you think spotting a rare earth deposit is tough, try identifying one when it looks exactly like barren dirt. That’s the challenge Dr Tobias Bamforth, Postdoctoral Research Fellow at Monash University, is helping explorers overcome.

Presenting “Defining Regolith-Hosted REE Deposits Using Multivariate Whole-Rock Geochemistry” at the AusIMM Critical Minerals Conference 2025 in Perth, Tobias and his co-authors from CSIRO, Monash University, Murdoch University, and OD6 Metals demonstrated how data science is transforming exploration of Western Australia’s regolith-hosted rare earth element (REE) deposits.

Seeing the unseen

Regolith-hosted REE systems are notoriously difficult to characterise. The mineralised and barren clays look the same under a microscope, and traditional tools like XRD can’t pick up the trace concentrations that matter. “You can be given the sample itself, the mineralogy, the grade and even the recovery, and still not tell whether it’s mineralised,” Tobias said during his presentation.

To tackle that problem, his team turned to multivariate statistics—specifically k-means clustering and principal component analysis (PCA)—applied to more than 3,000 whole-rock geochemical samples from the Splinter Rock REE prospect in Western Australia’s Esperance region. The goal was to separate transported sediments from in-situ weathered granite, and to map where rare earth enrichment actually occurs.

The algorithm finds the ore

The results were striking. The algorithms automatically grouped samples into five distinct stratigraphic horizons, forming laterally continuous layers that matched the expected regolith profile—three barren and two mineralised. “What’s exciting,” Tobias explained, “is that the data self-organised into the same layers we’d interpret geologically, purely based on chemistry.”

When the model was cross-checked against HyLogger hyperspectral data, the geochemical clusters aligned with known transitions between transported cover, saprolite, and saprock. The combination of machine learning and spectral validation gave explorers a new, scalable way to map subsurface boundaries without expensive and time-consuming mineralogical work.

More than numbers

The implications go beyond classification. By integrating whole-rock chemistry with mineralogical and metallurgical datasets, Tobias’s team identified several consistent trends across the Splinter Rock system:

• Highest REE grades occur within the granitic saprolite and saprock zones.

• Heavy REEs are enriched at the saprolite–saprock boundary, while light REEs decrease with depth.

• Magnet REEs—neodymium, praseodymium, dysprosium, and terbium—tend to accumulate in the saprock, where optimal metallurgical conditions also occur.

• Zones of enrichment correlate with negative cerium anomalies, a potential geochemical fingerprint for targeting new prospects.

As Tobias put it, “With this approach, we’re not just describing what’s in front of us—we’re predicting where the next mineralised horizon could be.”

Smarter sampling for smarter exploration

For explorers chasing clay-hosted and regolith-hosted rare earths, the workflow offers a fast, quantitative way to define sampling domains and focus drilling. Once the geochemical clusters are established, new samples can be classified instantly without repeating the full analysis.

“This makes it a living system,” Tobias said. “You can keep adding new data, and the model evolves with your understanding of the deposit.”

The approach also links directly to metallurgy. By knowing where the boundary between saprolite and saprock lies, explorers can target zones where leachability and acid consumption are optimised—an economic advantage for projects moving toward feasibility.

A digital toolkit for the clay REE age

Tobias’s work highlights a broader trend in mineral exploration: the convergence of geoscience, data analytics, and machine learning. Regolith-hosted REE deposits are low grade but high tonnage, and understanding their geometry and chemistry is key to unlocking their potential.

The combination of geochemical clustering, spectral validation, and mineralogical calibration offers explorers a data-driven framework for tackling the most elusive deposits in Australia’s critical minerals landscape.

In short: by teaching computers to read the regolith, Tobias Bamforth and his collaborators are giving explorers the ultimate map—one that reveals what the eye can’t see and the drill core can’t always tell.