How AI Cluster Detection Made XRT Ore Sorting Viable at Scale

At Hemerdon, roughly 70% of everything pulled from the ground never enters the main processing plant. That is not a shortfall or a bottleneck. It is a deliberate design choice, and it should make you look differently at every other processing operation that still grinds and mills the lot.

The traditional model processes first and discards later. It crushes, grinds, floats, and separates enormous volumes of rock, then throws away whatever fails to carry value. The sorting-first model inverts that logic entirely: it discards at the front and processes only what has been confirmed as worth processing. Making that work at industrial scale meant solving a specific technical problem in XRT ore sorting that had, until recently, kept the technology on the margins of high-throughput plants.

The traditional model processes first and discards later, running entire ore volumes through crushing, grinding, flotation, and jigging before any waste separation occurs; mineral processing at that scale carries enormous embedded energy cost in every tonne of barren rock that enters the circuit.

After reading this, you will understand precisely why OBTAIN’s particle-cluster detection changes the economics of that sorting-first model, and why it matters for processing low-grade, complex ores well beyond a single Devon tungsten mine.

Why traditional XRT sorting hit a ceiling in high-throughput operations

Push feed through a legacy sorter fast enough and something counterintuitive happens: the machine gets worse at its job. Throughput rises, belt occupancy climbs, and detection accuracy quietly falls away. That is the signal a plant operator sees, and the physics behind it is what kept XRT sorting out of the front end of high-volume plants for years.

XRT stands for X-ray transmission. The technology fires X-rays through particles moving on a belt and reads the differential transmission signal, using it to tell dense mineralised material apart from lighter waste rock. Denser, mineral-bearing fragments absorb more of the beam; barren rock lets more through. The machine ejects on that difference.

At Hemerdon, the feed is split by size before it ever reaches a sorter:

• Material below 10mm bypasses the sorters entirely, heading to pressure jigging and dense media separation circuits.
• Three COM Tertiary XRT machines handle the 10-30mm fraction.
• Three COM XRT 2.0 units process the 30-80mm fraction.

All six TOMRA XRT sorters run in parallel inside a dedicated sorting house at the front of the plant. Tungsten mineralisation here occurs primarily as wolframite inclusions locked within host rock, which is exactly why particle-level detection precision matters so much for recovery.

Why precision is non-negotiable at Hemerdon The valuable tungsten sits as wolframite inclusions inside the host rock rather than as clean, separable grains. Miss a particle, and you send recoverable tungsten to the waste pile.

The cluster problem explained

The trouble begins when the belt fills up. At high occupancy, particles stop travelling as neat, separated fragments and start touching, overlapping, and forming irregular combined shapes. The legacy algorithm reads that jumble as a single object.

The ejection decision that follows is binary and blunt. The whole cluster is either kept or rejected. Reject it, and any tungsten trapped inside goes to waste. Keep it, and waste rock rides forward to occupy expensive downstream processing capacity.

The practical result is systematic over-rejection of mixed waste-and-mineral clusters and systematic under-recovery of fine mineralised particles buried within them. For you, the takeaway is structural: before OBTAIN, plant designers faced a direct trade-off between throughput and recovery quality. Run the sorters faster and you lost product. That trade-off was the architectural constraint the entire sorting-first design had to defeat.

How OBTAIN’s convolutional neural network disaggregates the cluster

To see why OBTAIN matters, watch what it actually does with the belt image. Instead of treating a cluster as one lump, it works through the problem the way a careful human eye would, separating the crowd into individuals before deciding anything.

OBTAIN applies convolutional neural networks, a form of deep learning built to interpret images, to the raw XRT intensity data. Here is the sequence:

1. The system captures the XRT image across the full belt.
2. The convolutional network segments that image into individual particle regions.
3. It infers boundaries within clusters, drawing separations even where particles physically touch.
4. It assigns an independent ore-versus-waste probability score to each inferred region.
5. The sorter actuates precise ejection based on those individual scores, not on the cluster as a whole.

That fourth step is the pivot. By scoring each region independently, OBTAIN decouples the ejection decision from the cluster, breaking the old link between belt occupancy and detection accuracy.

The payoff is quantifiable. TOMRA reports that OBTAIN delivers approximately 80% higher feed per sorter while holding equivalent product quality and recovery rates. This is not a marginal tweak to an existing parameter. It is a structural change to the relationship between how full the belt is and how accurately the machine reads it, which is the technical reason the sorting-first plant became viable at Hemerdon’s scale.

Before finalising the plant design, Tungsten West conducted substantial bulk testing jointly with TOMRA Mining, validating sorter performance against Hemerdon’s specific feed characteristics. All six sorters are equipped with OBTAIN.

“The TOMRA technology package is a meaningful step-change for the project, expected to improve long-term production resilience,” said Jeffery Court, Chief Executive Officer of Tungsten West.

TOMRA’s related CONTAIN product, launched in June 2025, extends the same deep learning approach to a different challenge: detecting subsurface mineral inclusions in inclusion-type ores like tungsten. It is a signal the neural network toolkit is being applied across a widening set of detection problems.

For you as an investor, the mechanism is the reason Hemerdon can credibly target 500 t/h throughput across six sorters. It also tells you where to focus on execution risk: the approach has been validated in bulk testing, but full-ramp throughput at nameplate capacity is a different proof point, and one still ahead of the project.

What sorting first actually means for processing economics and carbon

Once you grasp the mechanism, the economics stop looking like projections and start looking like consequences. Sorting first is not a bolt-on efficiency. It reorders where value and cost sit in the entire plant.

The split is roughly 30/70. About 30% of incoming material, the mineralised fraction, goes to the downstream concentrator for tungsten and tin recovery. The remaining 70% exits as premium aggregate, bypassing energy-intensive grinding, milling, flotation, and jigging altogether.

That 70% becomes a genuine secondary cash stream. Life-of-mine aggregate production is projected at 45.8 Mt, ramping to 0.5 Mtpa later in mine life, with total premium sales reaching 10.9 Mt at an average 0.35 Mtpa until year 15. The base case models a price of £19/t ex-works, with offtake agreed with GRS Roadstone.

Aggregate leaves the site by three routes:

• Short road haul to nearby markets
• Rail from Marsh Mills
• Coastal shipping from Plymouth

The mix that drives project economics, though, is weighted heavily toward metal:

One caveat sits on the aggregate volume. Devon County Council planning conditions require aggregate to remain ancillary to tungsten and tin operations, capping site logistics at 50 lorry movements per day. That ceiling limits how far the aggregate stream can grow.

The carbon-boundary problem

The headline environmental claim is that the aggregate stream carries a near-zero incremental carbon footprint, because the 70% that becomes aggregate skips the energy-hungry processing stages. That is directionally true for downstream emissions, but it depends entirely on where you draw the system boundary.

A system boundary determines which emissions are attributed to the sorting operation and which are treated as pre-existing site emissions. Devon County Council regulators note that upstream crushing and sorting emissions occur whether material ends up as aggregate or as tailings. Those emissions do not disappear because a fragment was diverted early.

Environmental assessors point to a comparable installation for scale. An energy balance study at the Rossing mine found that introducing an ore sorter plant produced a 14-15% increase in annual greenhouse gas emissions, driven by auxiliary electricity, compressed-air demand, and reject rock transport.

That figure is not a verdict on Hemerdon. It is a reminder that the auxiliary energy demands of high-throughput sorting are real and should be modelled, not assumed away. If you are weighing the ESG case, the near-zero claim is boundary-sensitive: its validity rests on what the boundary includes, not simply on the fact that downstream processing has been eliminated for most of the rock.

Tungsten prices, strategic capital, and what the technology still needs to prove

The commercial backdrop for Hemerdon is, by any measure, extraordinary. The mine is reaching first production into a tungsten market priced far above anything its feasibility study contemplated.

The price gap driving Hemerdon’s economics Feasibility base case: US$400/mtu APT Early September 2026 market: US$2,900-3,100/mtu APT CIF Rotterdam That is roughly a 7.5 times uplift on the original model.

That gap is so wide that even partial delivery of the sorting plant’s throughput targets would still produce economics comfortably above the original project model. Government conviction has followed the numbers. In August 2026, the UK committed up to £71 million via the National Wealth Fund, comprising £36 million in equity and up to £35 million in debt, and securing rights to procure up to 50% of annual tungsten production.

The strategic logic is straightforward. Hemerdon represents roughly 4.1% of global tungsten production potential and is one of the largest single resources outside China. The UK Critical Minerals Strategy aims to limit single-country imports to no more than 60% of supply and source at least 10% of UK demand domestically by 2035, and the EU has classified the project as a strategic initiative under its own framework.

The UK Critical Minerals Strategy sets the policy framework that makes the National Wealth Fund commitment to Hemerdon coherent: it targets limiting single-country imports to no more than 60% of supply and sourcing at least 10% of UK demand domestically by 2035, with tungsten sitting explicitly among the priority minerals.

None of that resolves the live risks. Three sit at the front of the queue:

• Throughput execution: the 500 t/h target for the AI-enabled sorting plant has been validated in bulk testing but not yet demonstrated at full ramp.
• Financing: a US$25 million bridge facility at SOFR +4.5% remains in place, pending a larger facility of up to US$85 million to fully de-risk operations.
• Price sensitivity: elevated European APT prices stay highly sensitive to Chinese export policy, a structural market risk the sorting technology does nothing to mitigate.

The plant is designed to process 3.5 Mtpa of primary ore, targeting 332,000 mtu of WO3 annually, with full commissioning and ramp-up to nameplate capacity targeted for Q1 2027. Hold the price windfall against that one unresolved fact: 500 t/h remains a target, not yet a demonstrated operating rate.

What the Hemerdon model signals for the future of low-grade ore processing

Step back from Hemerdon and the sorting-first architecture starts to look less like a one-off and more like a response to pressures building across the critical minerals sector. Three of them are structural:

• Ore grades are declining, so more rock must be moved for the same metal.
• Energy costs are rising, making the grind-everything model steadily more expensive.
• The environmental licence to operate increasingly demands a smaller processing footprint.

Together, those pressures make the traditional approach, crushing and milling everything and sorting value out at the end, progressively harder to justify. Sorting first attacks the problem from the opposite direction.

OBTAIN’s resolution of the cluster problem is what makes that architecture usable at volume. By removing the throughput ceiling that once ruled XRT sorting out as a primary front-end tool, it broadens the range of high-volume, low-grade, hard-rock operations where the technology can credibly sit at the head of the plant rather than downstream.

Ore sorting at Pilgangoora provides a useful point of comparison: the P1000 lithium operation uses sensor-based pre-concentration at industrial throughput to reduce the volume of spodumene-bearing rock entering energy-intensive downstream circuits, applying the same sorting-first logic Hemerdon is deploying for tungsten recovery.

A widening toolkit TOMRA’s CONTAIN product, launched June 2025, applies deep learning to inclusion-type ore classification, evidence the neural network approach is being extended well beyond cluster detection.

The policy backdrop reinforces the direction. Both the UK Critical Minerals Strategy and the EU framework prioritise domestic processing capability, which favours plant designs that cut processing footprint and energy intensity.

Whether you are sizing up Hemerdon as an investment or tracking sensor-based sorting as a technology trend, the Q1 2027 ramp-up will tell you more than any feasibility projection. Seven TOMRA XRT units were delivered to the UK, six duty and one standby, and that redundancy reflects both confidence and the reality that this is a novel front-end configuration. It will be the first full-scale demonstration of OBTAIN-equipped XRT sorting as the primary front end of a producing hard-rock mine.

Where the sorting-first thesis stands as Hemerdon approaches full ramp

Separate what is settled from what is still contingent, and the picture becomes usable. Three things the deployment has already demonstrated, and one it has not:

1. Bulk-tested, OBTAIN-equipped XRT sorting can be engineered as a primary front end at industrial scale.
2. The sorting-first design produces a commercially viable aggregate by-product stream.
3. The strategic capital and policy frameworks are in place, confirmed by the £71 million National Wealth Fund commitment in August 2026, to carry the project through ramp-up.
4. Still outstanding: whether the 500 t/h throughput target holds at nameplate capacity with the product quality and recovery rates seen in bulk testing.

With six duty sorters and one standby delivered, and APT sitting at US$2,900-3,100/mtu in early September 2026, the ingredients are assembled. Q1 2027 is when the technology argument moves from engineered potential to demonstrated performance.

That distinction, between bulk-test validation and full-ramp production, is the most valuable single piece of context you can hold on this project. It lets you read commissioning updates with precision rather than reacting to headline production numbers without the processing context to interpret them.

For readers tracking this space across multiple commodities and geographies, our dedicated guide to sensor-based sorting technologies covers how XRT, laser, and near-infrared modalities are being deployed across hard-rock, industrial mineral, and battery metals operations, with analysis of throughput benchmarks and capital payback profiles.

This article is for informational purposes only and should not be considered financial advice. Investors should conduct their own research and consult with financial professionals before making investment decisions. Past performance does not guarantee future results. Financial projections are subject to market conditions and various risk factors, and forward-looking statements are speculative and subject to change based on market developments and project performance.