Mining & metals · Case study
Same ore. Same plant. More Copper.
A production data and ML system across the full circuit, from ore feed to final concentrate, holding recovery at its ceiling as ore and plant conditions shift.
Ore → Grind → Flotation → Concentrate
Full sensor-to-warehouse data backbone
We deliver the downstream data and ML: the backbone, the models and the optimiser. Copper is live; India and South-East Asia metals are next.
Recovery left behind at any stage is gone for good. Precise, coordinated setpoints at every stage are what capture it, which is why model accuracy at every stage compounds.
Case study · Copper recovery, data science and ML
In production
Holding copper recovery at its ceiling, mine to concentrate.
Major copper concentrator, Indonesia
The problem
Copper recovery is won or lost stage by stage: feed grade, grind size, reagent dosage, flotation setpoints. The best settings move constantly as ore hardness and feed quality change, and no operating team can re-tune the whole circuit by hand fast enough to keep up. Recovery left behind at any stage is gone for good, and at this scale every fraction of a percent is worth millions a year.
What we built
A production data-science and ML system across the circuit. Every process variable, from grade and tonnage to power, reagent dosage and particle size, streams from plant sensors into a single analytical source of truth in near real time. On top of it sit 31 production models: 8 for feed characterisation, 4 for grinding and throughput, 4 for recovery and grade, 7 for final quality, and 8 XRF assay models that read S, Fe and Cu chemistry from spectra. Each predicts a stage outcome and recommends the setpoints that move it toward maximum recovery.
Plant-wide, not stage by stage. The optimiser coordinates flotation setpoints across the whole train, so a gain in one cell never quietly costs recovery downstream.
Assays in near real time. Eight XRF models read S, Fe and Cu chemistry from spectra, without waiting on lab results.
Self-sustaining. Models are watched for drift and retrained automatically under human sign-off, without a data-science team babysitting them.
Business outcomes
In production
31 ML models
Deployed from feed characterisation through final concentrate quality.
In production
Near real time
A sensor-to-warehouse data backbone across the mine-to-concentrate chain.
In production
Plant-wide
A flotation optimiser producing setpoint recommendations operators can act on.
In production
Self-retraining
Models watched for drift and retrained automatically under human sign-off.
Value at stake
≈$20M a year
Value at stake for every +0.5% recovery at this plant’s throughput.
Why it holds
Model accuracy at every stage compounds: coordinated setpoints are what capture the recovery tuning by hand leaves behind. On the roadmap: ore-regime-conditioned models, and a full mining optimiser that extends the coordinated objective upstream through crushing and grinding, closing the loop from mine face to final concentrate.
Stack
31 production ML models · XRF assay models · Plant-wide flotation optimiser · Drift monitoring and retraining