Ore sorting isn’t a new concept. The idea of separating valuable material from waste before it reaches the processing plant has been around for decades. What’s new is the combination of advanced sensors and machine learning that’s making it work reliably at production scale — and the results are starting to shift how mining companies think about their entire value chain.

The Basic Proposition

Every mine sends a certain amount of waste rock through its processing plant. In a typical gold operation, you might be processing ore at 2-3 grams per tonne, meaning 99.99% of what goes through the mill is effectively waste. Even in higher-grade base metal operations, significant volumes of sub-economic material make it to the plant, consuming energy, grinding media, reagents, and water to process rock that doesn’t pay for itself.

Ore sorting aims to remove that sub-economic material earlier in the process — ideally right at the mine face or on the ROM pad — before it consumes processing capacity.

The economic case is straightforward: if you can reject 20-30% of mined material as waste before it reaches the plant, you’ve effectively increased your plant’s capacity by that same percentage without building additional infrastructure. Or you can maintain the same throughput with a higher average head grade, which means better metal recovery and lower unit costs.

Where AI Fits In

Traditional ore sorting used simple threshold-based rules. If a rock’s X-ray fluorescence reading exceeds a certain value, keep it. Below the threshold, reject it. This works for binary situations — you’ve got ore or you don’t.

Real geology isn’t that clean. Ore bodies are heterogeneous. Mineralisation is complex. A single threshold might reject material that’s actually economic under certain conditions, or accept material that’s not worth processing.

Machine learning changes this by building classification models trained on thousands or millions of sensor readings correlated with actual assay data. These models can identify subtle patterns in sensor responses that correspond to different mineralisation types, alteration assemblages, and grade distributions.

Companies like TOMRA Mining and Steinert have been leading the sensor hardware side, while the AI layer is increasingly being developed by specialist firms or in collaboration with university research groups.

The Team400 approach to this kind of industrial AI — building custom models trained on operation-specific data rather than generic solutions — is particularly relevant here. Every ore body is different, and a sorting model trained on one deposit won’t necessarily transfer to another without significant retraining.

What’s Actually Working in 2026

Several commodity types are seeing strong adoption:

Diamonds. The diamond industry was an early adopter, and X-ray transmission sorting has been standard practice at major producers like De Beers for years. AI-enhanced systems are now improving recovery rates and reducing false positives.

Gold. Sensor-based sorting using dual-energy X-ray transmission (DE-XRT) and near-infrared (NIR) sensors is proving effective at several operations. The key challenge with gold is that it’s often present in such fine particles that bulk sorting (rejecting whole rocks based on surface or transmission measurements) has limitations. But for removing obviously barren material — like un-mineralised waste dilution from stoping — it works well.

Base metals (copper, zinc, nickel). This is where some of the most exciting developments are happening. XRF sensors can directly measure metal content on rock surfaces, and when combined with AI classification, they can sort copper ore from waste with impressive accuracy. First Quantum Minerals has published results showing significant waste rejection rates at their Kansanshi operation.

Iron ore. Sorting by grade is relatively straightforward with iron ore, and several Australian iron ore producers are actively trialling or deploying sensor-based sorting systems to upgrade run-of-mine material before it reaches the beneficiation plant.

The Infrastructure Challenge

The technology works. The harder part is fitting it into existing operations.

Ore sorting works best on sized material — typically in the 20-200mm range. That means you need a crushing and screening step before the sorter, which adds capital cost and another unit operation to manage. At a greenfield operation, you can design this into the flowsheet from the start. At an existing operation, finding space on the ROM pad and integrating a sorting circuit into existing material handling can be a logistical headache.

Throughput is another consideration. Current generation ore sorters can handle hundreds of tonnes per hour, but large operations processing 50,000+ tonnes per day may need multiple sorter banks running in parallel. The capital cost of the sorting circuit needs to be weighed against the savings in downstream processing.

Water is worth mentioning too. Ore sorting is essentially a dry process, which makes it attractive in water-scarce regions. Unlike gravity or flotation-based pre-concentration, you don’t need water to make it work. In Australia’s arid mining regions and parts of South America and Africa where water is a constrained resource, this is a meaningful advantage.

Grade Engineering and Mine Planning

Where this gets really interesting is when ore sorting data feeds back into mine planning. If you know that your sorters can reliably reject waste at a certain efficiency, you can change your mining strategy. You might choose to mine at a lower cut-off grade, knowing that the sorter will upgrade the feed. That potentially extends mine life and opens up material that was previously considered sub-economic.

This concept — sometimes called “grade engineering” — was championed by CRC ORE in Australia and is gaining traction globally. The idea is that you think about grade management across the entire value chain, from blast design through to processing, rather than just relying on geological grade control.

The Bottom Line

Ore sorting with AI classification isn’t speculative anymore. The hardware is mature, the algorithms are proven, and the economic case is strong for many applications. The operations adopting it now are seeing reduced processing costs, improved head grades, lower water and energy consumption per tonne of metal produced, and extended mine lives.

For operations still evaluating, the question isn’t really whether ore sorting works. It’s whether the specific ore type, mineralogy, and operational context make it viable. That’s a testwork question, and the path to answering it — bench-scale sensor testing followed by pilot-scale trials — is well established.

The mines that figure this out early will have a structural cost advantage that compounds over the life of the operation. That’s not hype. That’s just physics and economics.