At the UK's largest tungsten mine, an AI decides particle by particle what is worth processing

Tungsten West installed six AI-powered X-ray sorters at the Hemerdon mine to reject 70% of run-of-mine ore before the processing stage — cutting cost, energy and the size of the plant required.

In September 2026, Tungsten West restarted the Hemerdon tungsten mine in the English county of Devon with a processing plant redesigned around six X-ray transmission (XRT) sorters made by TOMRA Mining. All six machines use the deep learning technology called OBTAIN, which TOMRA developed to identify the properties of individual particles at high speed. Material leaving the crushing circuit, between 10mm and 80mm, is split into three fractions: anything below 10mm goes to pressure jig circuits; the 10mm to 30mm fraction passes through three COM Tertiary XRT machines; and the 30mm to 80mm fraction goes through three COM XRT 2.0 units. The result of the sorting is that roughly 30% of the material — the mineralized fraction — goes on to the concentrator, where tungsten and tin are recovered, while the other 70% is rejected as premium-quality construction aggregate, with an additional carbon footprint close to zero. The company expects to ramp up to full-scale production in the first quarter of 2027.

X-ray sorters are nothing new in mining — the technology uses the transmission of X-rays through rock to distinguish atomic composition and density, separating ore from waste before any more expensive chemical or physical processing. The historic problem with this kind of sensor is that it sees material at high speed on a belt, and particles stuck to each other — clusters of small rock attached to larger fragments — tend to be read as a single object, which leads the system to accept or reject the whole block based on an imprecise reading. That blind spot is exactly what OBTAIN tackles: the neural network algorithmically breaks down each cluster detected by the sensor into individual particles, allowing the sorter to make an accept-or-reject decision per particle, not per block. In practice, according to TOMRA, this allows each sorter to be fed about 80% more material while keeping the same quality and the same recovery rate — a capacity gain that would normally require buying more equipment, not just smarter software running on what already exists. The economic effect of rejecting 70% of run-of-mine ore mass before the concentration stage is bigger than it looks at first glance: every tonne that does not go through the concentrator is one tonne less consuming water, energy and reagent at the most expensive stage of processing. And the rejected material does not become discarded tailings — it is sold as construction aggregate, which turns a disposal cost into additional revenue.

Hemerdon has tried to operate before and did not survive economically the first time, which makes process engineering, not geology, the deciding factor in this second attempt. Sensor-based sorting ahead of the most expensive chemical or physical stage is one of the few mining technologies that simultaneously reduces capital cost, energy consumption and tailings volume — and that is what can make the difference between a marginal deposit staying uneconomic or becoming commercial production. The timing is not neutral either: tungsten is on the UK's critical minerals list, China controls about 80% of global production, and Chinese export restrictions have been reshaping the global supply chain — with demand projected to grow 2% to 5% a year through 2033. Hemerdon, operating at full scale, would be able to supply about 4.1% of global tungsten production, a small share in absolute terms, but relevant as a non-Chinese source of a mineral that today has few supply alternatives outside China.