Reviewed August 2026 against MSHA fatality and open-data reporting, the USGS–DARPA CriticalMAAS programme, and Freeport-McMoRan’s filed results.

Try it: Recovery-Uplift Value Calculator →

AI in Mining: What Predictive Systems Actually Deliver

AI in mining is four separate jobs wearing one label: models that rank exploration ground before a rig moves, process-control models that lift recovery inside an existing plant, condition models that call an equipment failure early, and predictive monitoring that infers an emission or a slope movement from process data instead of a direct sensor. Only one of those four has produced publicly reported, SEC-filed metal at scale — Freeport-McMoRan booked 214 million pounds of incremental copper in 2025 from its leaching and technology initiatives. The rest is either a funded government programme with published milestones, or a vendor claim you should treat as unproven until it survives the seven-question checklist further down this page.

This article gives you the figures that exist, names who published them and when, and shows you the public database where you can pull a fresher number than the one printed here. Where a number does not exist in public, it says so rather than inventing one.

Predictive Mining, Defined Precisely

“Predictive mining” is not one product. Buyers get burned because a vendor selling vibration analytics and a vendor selling prospectivity maps both answer to the same phrase. The table below separates them by what each actually predicts, what it needs as input, and — the part most marketing omits — how you would independently check whether it worked.

Table: The four layers sold as “predictive mining”, and the test each one has to pass before you believe it.

The stakes are not abstract. MSHA’s rolling fatality report, pulled on 6 August 2026, showed 19 mining fatalities year-to-date — 6 underground and 13 at surface operations — against 17 at the same point in 2025, 14 in 2024, 28 in 2023 and 18 in 2022. That page updates as reports are filed, so the count you see will differ from the one below.

Fatality counts move around because the denominator is small; a single powered-haulage incident swings the line. That is exactly why safety cases for autonomy are argued on exposure hours removed rather than on year-over-year death counts. Check the live figure at MSHA’s daily fatality report.

Layer 1: Exploration Targeting — the Best-Documented Public Case

The clearest public evidence that machine learning changes exploration economics comes from the US government, not from industry. DARPA’s Information Innovation Office and the USGS ran CriticalMAAS (Critical Mineral Assessment with AI Support), launched in February 2024. A conventional USGS mineral resource assessment takes about two years start to finish. At the programme’s final hackathon in Reston, Virginia on 13–17 January 2025, the AI-assisted workflow completed an assessment in roughly two and a half days.

Two caveats the programme itself is honest about. First, a compressed workflow is not the same as a validated prediction — speed and hit rate are different metrics. Second, the pilots were assessments, not drill programmes. Between February 2024 and the close of the hackathon series, about a dozen pilot assessments were run across commodities including zinc, copper, nickel, cobalt, lithium, tungsten and rare earth elements. Assessing 50 commodities across 100 deposit types nationally by conventional methods would take many years; that gap is the entire argument for the technology. Read the programme write-up at DARPA.

None of it works without input data, which is where the second public programme matters. The USGS Earth Mapping Resources Initiative (Earth MRI) received $9.6 million in FY2019 and $10.6 million in FY2020, then $320 million under the Infrastructure Investment and Jobs Act of 2021 — $64 million per year for FY2022 through FY2026. Between 2019 and 2023 it tripled national coverage of high-resolution geophysical surveys, working with 40 states and territories, and the IIJA requires a comprehensive national surface and subsurface map by 15 November 2031.

Full detail is in the Congressional Research Service brief on Earth MRI. The practical read for a US explorer: the free public geophysical layer under your ground is denser than it was, and it is the same layer commercial prospectivity models train on.

Layer 2: Recovery and Process Control — Where the Audited Tonnes Are

This is the layer that shows up in financial statements. Freeport-McMoRan, the largest US copper producer, reports incremental copper production from leaching and technology initiatives that apply data analytics to existing stockpiles and plants across its US and South America operations. The company reported 214 million pounds of incremental copper for full-year 2025, an annualised run rate of approximately 240 million pounds reached in late 2025, 54 million pounds in the first quarter of 2026, and a target of approximately 300 million pounds for 2026.

Read that carefully. It is a company-reported operational figure covering leaching chemistry, applied heat, an internally developed additive tested at Morenci, and analytics together — not a clean attribution to a machine-learning model alone. That is normal and it is the honest way to report it, but it means you cannot quote “AI added 214 million pounds”. The source is the Freeport fourth-quarter and year-ended 2025 results release, dated 22 January 2026; quarterly updates land in late April, July and October.

Scale that logic to your own plant with the figures you actually have — not ours.

Layer 3: Predictive Maintenance and Autonomy

Predictive maintenance is the oldest of the four layers and the one with the weakest public evidence base in mining specifically. The most-cited authority is not a mining study at all: the US Department of Energy's Federal Energy Management Program O&M Best Practices guide, catalogued as PNNL-SA-37830, which frames operations and maintenance improvements as worth 5% to 20% on energy bills without significant capital investment. The record is public at OSTI, the DOE research repository. Vendors routinely quote 30–45% downtime reductions; those figures do not trace to a peer-reviewed mining dataset, and you should ask any supplier quoting them for the underlying site and period.

Autonomy is easier to see. Rio Tinto's Gudai-Darri iron ore mine in the Pilbara, which achieved first ore in 2022 and an annual average capacity of 43 million tonnes, runs autonomous haul trucks, autonomous drills and the AutoHaul rail network from an operations centre in Perth roughly 1,500 kilometres away. Its water carts carry on-board AI that monitors dust and sprays automatically, on a 160,000-litre tank — a 33% increase on the operation's previous largest carts. Details are in Rio Tinto's own account of the site, published 12 August 2022.

Layer 4: Predictive Monitoring, Including Emissions

Predictive monitoring means inferring a quantity you are not directly measuring from parameters you are. In US environmental compliance this has a formal legal shape. A predictive emission monitoring system (PEMS) uses process and control-device operating parameters plus a conversion equation or model to produce a result in the units of the applicable emission limit — an alternative to a hardware continuous emissions monitoring system. EPA's specification for certifying one, Performance Specification 16, "Specifications and Test Procedures for Predictive Emission Monitoring Systems in Stationary Sources", sits in Appendix B to 40 CFR Part 60; you can confirm it is listed at Cornell's Legal Information Institute.

Why a mining audience should care: mine-site power generation, smelters, kilns and dryers are permitted stationary sources. Where a reliable correlation exists between operating conditions and emissions, PEMS can be petitioned as an alternative monitoring system — the certification and quality-assurance procedures are found under 40 CFR Part 60 Appendix B and Part 75 Subpart E. No federal rule mandates PEMS; approvals are source-specific petitions. That is a meaningful distinction: a predictive emissions model is a regulated instrument with a relative-accuracy test to pass, not a dashboard.

The same discipline is worth importing into unregulated predictive monitoring — tailings dam piezometer trends, InSAR-derived slope movement, haul-road dust. If a model would not survive a relative-accuracy audit against a reference method, it is a decision aid, not a monitor, and it should not be the last line of defence for anything.

The Seven-Question Checklist for Any Predictive Mining Claim

This is the durable part of this article. The figures above will age; these questions will not.

1. What exactly is being predicted, in what units, over what horizon? "Better recovery" is not a prediction. "Flotation recovery for the next 8-hour shift, in percentage points, ±1.2" is.
2. What was the baseline, measured over what period? A result quoted against a bad quarter is not a result.
3. Was the model tested on data it never saw? Ask for out-of-sample or forward-test performance with dates, not a fit to history.
4. What is the false-positive rate? For maintenance and safety models this determines whether crews keep using it.
5. Who else could have caused the improvement? Freeport reports leaching chemistry, applied heat, an additive and analytics as one programme — a candid disclosure most vendors avoid.
6. What happens when the ore changes? Models trained on one ore domain degrade when the pit moves. Ask for the retraining trigger and cadence.
7. Can an outsider reproduce the claim from public data? Production, employment and injury figures for any US mine are filed with MSHA and downloadable. If a claim cannot be cross-checked there, treat it as marketing.

Copper: Why the Capital Is Chasing Prediction

Copper is where predictive mining gets funded first, because grades decline while demand for grid and electrification copper rises, and because a fraction of a percentage point of recovery on a large concentrator is worth more than most capital projects. For US context, the copper chapter of the USGS Mineral Commodity Summaries — published each January by the National Minerals Information Center — gives recoverable copper from US mine production, its value, the Arizona share of domestic output, and net import reliance, all on a consistent annual basis. It is the reference to quote in any investment memo, and it supersedes any figure on this page.

Price is the other half of the equation and the half nobody should hard-code. If you are modelling what a recovery gain is worth, use your own realised price from your offtake terms, not a spot headline. For a structured view of the forecasting arguments, see our separate analysis of copper price prediction.

Farmonaut: Satellite Mineral Intelligence for Predictive Mining

Farmonaut works the first layer — targeting — and is deliberately explicit about what that does and does not include. Our satellite-based mineral detection service analyses multispectral and hyperspectral Earth observation imagery with machine learning to map alteration mineralogy and surface spectral signatures across a licence area, producing ranked target zones before a rig is mobilised.

• 🛰️ What it measures: surface and near-surface spectral response — alteration halos, iron oxides, clays, carbonates — not metal at depth.
• 🔬 What it outputs: ranked prospectivity zones with GIS-ready files you can load beside your Earth MRI geophysics.
• 🌎 Where it has run: projects across Africa, South America, Australia and North America, spanning more than 80,000 hectares, 18+ countries and 13+ mineral targets.
• ⏱ Turnaround: most reports are delivered in 5–20 business days from receipt of coordinates and target minerals.
• 🌿 Ground impact: zero — no access road, no camp, no disturbance to screen a district.

Premium+ reporting adds interactive 3D subsurface modelling for vein orientation and drill targeting, GIS-ready exports, and commercial conclusions for investment decision support. A worked example of satellite driven 3D mineral prospectivity mapping shows how drilling routes are planned from it. The honest framing: this reduces the ground you drill, it does not replace drilling, and it should be scored on hit rate against your own historical baseline — question 1 of the checklist, applied to us.

How to Verify Everything on This Page

MSHA publishes 25 open datasets covering every coal and metal/non-metal mine in the United States since 1970, with accident, injury and illness records from 2000 onward, quarterly and annual employment and production files, inspections, violations and controller history. The files are refreshed every Friday afternoon. Start at MSHA's open government data page, download the Mines and Employment/Production sets, and join them on mine ID. That single join lets you check any operator's reported employment and production trend against whatever a technology case study claims — which is the only defence against unverifiable numbers, including ours.

Frequently Asked Questions

Is AI in mining actually producing metal, or is it still pilots?

Both. Process and leach optimisation is producing reported metal — Freeport reported 214 million pounds of incremental copper in 2025 from leaching and technology initiatives and targets about 300 million pounds in 2026. Exploration AI is at the validated-pilot stage: the USGS–DARPA CriticalMAAS programme ran roughly a dozen pilot assessments from February 2024 and demonstrated a 2.5-day assessment workflow against a conventional two-year timeline.

What is a predictive emission monitoring system?

A PEMS predicts an emission rate from process operating parameters via a model, instead of measuring the stack gas directly. EPA's Performance Specification 16 in Appendix B to 40 CFR Part 60 sets the certification test. No federal rule requires PEMS; sources petition to use one in place of a CEMS, and approvals are granted case by case.

How much does predictive maintenance save at a mine?

There is no reliable public, mining-specific figure, and any article quoting one without a named site and period is guessing. The nearest defensible authority is the DOE FEMP O&M Best Practices guide (PNNL-SA-37830), which frames O&M improvement as worth 5–20% on energy bills without significant capital. To get a real number for your operation, baseline unplanned downtime hours per major asset over four consecutive quarters before deployment.

Does satellite mineral detection replace drilling?

No. It measures surface spectral response and ranks ground; drilling is what confirms grade and continuity. Its value is in the ratio of holes drilled to discoveries made, which you can only judge against your own prior hit rate.

Which figures on this page will go stale first?

The MSHA fatality count changes continuously; Freeport's leach figures update quarterly, in late April, July, October and January; Earth MRI's IIJA funding line runs through FY2026 and then depends on new appropriations. The checklist and the MSHA verification method will still work when all of those numbers have moved.

Resources & Contact

• 📄 Request a mining intelligence report: Get a quote
• 📬 Speak to the mining analytics team: Contact us
• 🗺 Map a site: mining.farmonaut.com
• 🛰 Service detail: Satellite-based mineral detection
• 🧭 Sample deliverable: 3D prospectivity mapping example

Predictive mining is real where it is measured and mythical where it is not. Pick one layer, set a baseline you can defend, and hold the model to the same standard EPA holds a predictive emissions monitor to — a documented accuracy test against a reference. Everything else is a slide deck.