At Freeport-McMoRan’s Bagdad mine in Arizona, AI analysis helped operators reconsider a single set of mill settings. McKinsey, which worked on the project, described seven ore groupings with different processing responses in its February 2020 account.[1]
Separating copper minerals from the rock
The University of Arizona explains the physical sequence for sulphide ores. Crushing and grinding reduce the rock to small particles, which are mixed with water. In flotation tanks, reagents help copper-bearing mineral particles attach to rising air bubbles. The resulting froth is collected for further processing.[3]
This produces a concentrate that still needs thickening, smelting and refining. A mill’s output is therefore an intermediate material. Increasing its copper recovery means retaining more of the copper entering that stage, rather than making the incoming rock richer.[3]
Matching settings to the incoming material
Bagdad’s model combined thousands of decision trees through extreme gradient boosting, with metallurgical knowledge built into the analysis. It examined settings including feed characteristics and flotation-cell pH. The team used its recommendations to test operating changes; the process included the people running the mill.[1]
McKinsey’s June 2023 account reports a quarter above 85,000 metric tonnes of ore processed daily, 10% higher than the preceding quarter, alongside a one-percentage-point increase in copper recovery. These are the project participant’s operational results, not an independent trial isolating AI’s contribution.[2]
Why one percentage point can matter
The arithmetic has three parts: ore treated, copper contained in that ore and the fraction recovered. A hypothetical mill treating 80,000 tonnes a day at 0.6% copper receives 480 tonnes of contained copper. Holding those inputs fixed makes the effect of recovery easy to see.
Illustrative copper mass balance
| Recovery | Copper recovered daily | Copper remaining outside the recovered product |
|---|---|---|
| 85% | 408 tonnes | 72 tonnesSource 3 |
| 86% | 412.8 tonnes | 67.2 tonnesSource 3 |
Throughput can change independently. Processing more ore is useful only alongside knowledge of its grade and the resulting recovery. Those three measurements distinguish a faster mill from one that extracts more copper from each tonne it receives.
Testing a plant change before applying it
BHP’s January 2026 account describes a related approach at Escondida in Chile. AI-supported digital models combine live and historical information so operators can explore how ore characteristics or equipment settings might affect the plant. They can examine a proposed change virtually before applying it.[4]
BHP identifies natural variation, particularly ore type and hardness, as an operating challenge. Its account confirms practical use of these models but provides no recovery percentage for this specific application. That makes it useful evidence of the workflow, alongside the separately reported Bagdad results.[4]
The attraction is a processing plan that responds to the material actually arriving. A learned relationship becomes valuable when it suggests an actionable setting, and the operating record shows what happened after the team used it.
Sources & further reading
- Inside a mining company’s AI transformation McKinsey · 5 Feb 2020 · Accessed 12 Oct 2026
- Freeport-McMoRan turns data into value McKinsey · 22 Jun 2023 · Accessed 12 Oct 2026
- Copper Mining and Processing: Processing Copper Ores University of Arizona Superfund Research Center · Accessed 12 Oct 2026
- AI is improving performance across global mining operations BHP · 30 Jan 2026 · Accessed 12 Oct 2026