AI in Mineral Exploration: 6 Uses, Tools and Results

How AI is used in mineral exploration: six jobs

In mineral exploration, AI mostly means machine-learning models that rank ground for follow-up work. A model learns from maps, satellite scenes, geophysical surveys and the records of known deposits, then scores every cell of a map by how closely it resembles ground that already holds metal. The output is a ranked target list for a geologist to check. It is not a discovery. Where it works, it shortens the slow desk stage: a US Geological Survey (USGS) geologist told DARPA that a typical mineral resource assessment takes two years from start to finish, while hackathon runs of the AI-supported CriticalMAAS workflow brought it to two and a half days (DARPA, 15 January 2025). The same article states that AI-generated assessments still have to meet USGS scientific standards and peer review.

Chart

Days to complete one critical-mineral assessment

Typical USGS assessment (two years, shown as 730 days) against the AI-supported CriticalMAAS hackathon workflow

Six jobs appear again and again in the public work reviewed for this page. The table gives the data each one needs, what comes out and one dated example.

The Smackover figure shows why the wording matters. The authors report that the range equals 36 to 136% of the current US lithium resource estimate, yet it is a model output built from brine samples. It is not a drilled resource.

Satellite data in the models: which sensors map what

Most AI work that starts from space uses optical sensors. These measure reflected sunlight in many narrow slices of colour called spectral bands. The shortwave infrared (SWIR) bands matter most for mineral work, because clay and some carbonate minerals leave their marks there. NASA’s Landsat guide notes that SWIR bands help distinguish hydrothermal clay minerals for mineral exploration (NASA).

Chart

Spectral bands recorded by four satellite sensors

Total bands per sensor, including visible, infrared and (where fitted) thermal bands

The sensors differ in band count. Landsat 8 and 9 have 11 bands, with two main SWIR bands at 30 m. Sentinel-2 has 13, ten visible and near-infrared plus three SWIR (Copernicus SentiWiki). ASTER has 14, with six SWIR bands at 30 m, although the SWIR detectors have been unusable since April 2008, so only older scenes carry SWIR data (NASA JPL; NASA Earthdata). The German EnMAP satellite, launched on 1 April 2022, records 232 bands from 420 to 2450 nm at 30 m (DLR). That is hyperspectral imaging: so many narrow bands that each pixel holds something close to a full spectrum.

More bands do not mean deeper vision. All of these sensors see only the top of the ground: bare rock, soil or plant cover. A model trained on them finds alteration where rock is exposed and says little under thick cover. For the full sensor-by-sensor method, read our guide to remote sensing in mineral exploration. For what a satellite screen of a licence looks like as a product, see our page on satellite mineral detection, and for the alteration signals themselves, alteration mineral satellite mapping.

Machine learning mineral exploration in Canada: what public work shows

Canada’s federal effort sits in the Critical Minerals Geoscience and Data Initiative, led by the Geological Survey of Canada. Its stated short-term aims include applying artificial intelligence and machine learning to build national-scale models of mineral potential. The programme page lists models for carbonatite-hosted rare earth elements and for lithium in pegmatites; the lithium work uses generative adversarial networks, natural language processing and convolutional neural networks (Government of Canada). The page says the carbonatite models cut the search area by 80% while still predicting all known occurrences. It gives no method for that figure, so treat it as the programme’s own result. For background on the metals themselves, see our pieces on rare earth metals and critical minerals and what makes a metal a rare earth.

Chart

Exploration spending by company type

C$ billion, exploration and deposit appraisal, Canada

Chart

Canadian exploration and deposit appraisal spending

C$ billion; 2025 is preliminary and 2026 is company intentions

Money is moving the same way. Natural Resources Canada reports exploration and deposit appraisal spending of C$4.4 billion in 2025 (preliminary), after C$4.2 billion in 2024, with companies’ intentions for 2026 at C$5.3 billion, up 21% (NRCan bulletin, April 2026). Junior companies planned the bigger rise.

For a claim holder, the practical point is data. Models learn from assessment reports, geophysical grids and drill logs, and Canadian provinces publish many of them. Our guide to the Ontario Geological Survey, MINFILE and SIGÉOM shows where to find them before staking, and our Canadian exploration spending breakdown gives the province-level view. Readers following listed explorers can start from our list of Canadian gold mining companies.

Real results so far: four public examples

Four examples show what has been claimed and what is on record. Each row states who says it.

The pattern is the same in all four rows. The model narrows the search and the drill bit confirms or rejects it. Mingomba is the best-known case, but the same news report notes that construction began before engineering studies were complete, and the source does not say how much of the schedule came from AI.

AI tools for early-stage exploration compared

Searches for the best AI for geology or the best satellite-based exploration software have no single answer, because the tools do different jobs at different stages. The table sorts them by what you are trying to do. It names examples from the sources above and says what each source shows; it is not a ranking of vendors.

On the satellite row, Farmonaut states that early exploration timelines fall from months to days and that early-exploration cost is up to 80 to 85% lower than with ground-first methods. Its service covers gold, silver, copper, lithium, uranium, rare earth elements and more than 20 mineral types in all. Those are company figures, not independent test results.

Hiring mineral system modelling and mapping services

Mineral system modelling treats a deposit as the end of a chain: a metal source, a pathway for fluids, a trap and a place that survived erosion. A good service maps evidence for each link and then combines the maps. Mining mapping services range from geological and alteration maps to prospectivity models. Before you hire anyone, ask five questions:

• Which input datasets will be used, and who supplies them?
• How is the model tested? SRK Peru used spatial cross-validation based on geographic clustering, because random splits flatter a model when deposits cluster.
• What does the output look like for a geologist: probability classes, uncertainty, or one score?
• Which file formats come back, such as GeoTIFF or shapefile, so your team can load them in GIS?
• Who owns the model and the maps afterwards?

The SRK study is a good reason to ask the second question. It scored four models by area under the ROC curve (AUC), where 1.0 is perfect and 0.5 is a coin toss.

The highest AUC did not give the best map. SRK reports that the graph-based model produced results with greater spatial and geological coherence, while some other models tended to overfit. A score is a starting point for a geologist’s judgement, not a replacement for it.

▶ Six kinds of AI geology tools, what each does and three questions to ask before choosing one. · 3:18

Where AI falls short: labels, cover and ground truth

Deposits are rare, and a model needs examples. SRK describes its positive data points as limited and imbalanced. A 2024 study of porphyry copper in the Americas, which trained on known deposit locations only, found that performance in North America was high while scores for South America were relatively poor, and warned that care is needed when applying the method to new areas (GSA Bulletin, 2024). A model that works in one belt can fail in the next.

Three more limits apply. First, data quality: the USGS Smackover team found that three of 544 Gulf Coast brine samples read above 500 mg/L, and the source records could not confirm them. Second, cover: optical satellites see the surface, so soil, vegetation and transported sand hide the rock below. Third, proof: only drilling and assays turn a target into a resource, and a resource statement under NI 43-101 or JORC needs a qualified person (NI 43-101) or competent person (JORC). Satellite and AI outputs are exploration targets that sit ahead of that step.

▶ What AI in mineral exploration really does, where it fails and how to check any claim. · 4:37

How to use AI on your own licence: five steps and a budget check

For a licence or claim owner, a sensible order of work looks like this:

Infographic

AI screening of a licence in five steps

1. 1Fix the questionOne mineral, one deposit type, one boundary.
2. 2Collect existing dataGovernment geology, geophysics, assessment reports and satellite scenes.
3. 3Build evidence layersAlteration, faults and fractures, magnetic or radiometric grids.
4. 4Rank targetsRun a model, then have a geologist review the top of the list.
5. 5Check on the groundMap and sample the best targets; drill only what survives.

1. Fix the question: one mineral, one deposit type, one boundary.
2. Collect what already exists: government geology, geophysics, assessment reports and satellite scenes.
3. Build evidence layers: alteration from satellite data, faults and fractures, and any magnetic or radiometric grids.
4. Rank targets with a model, then have a geologist review the top of the list against field knowledge.
5. Check the best targets on the ground with mapping and sampling, and drill only what survives that check.

The calculator below tests the budget logic of step 4. It shows how much drilling money a ranking could save if the targets it drops would really have been barren. It cannot know that, so use it to see what a ranking must achieve to pay for itself, not to forecast a result.

▶ How a satellite picture becomes a tonnage and grade, and which inputs are measured or assumed. · 2:54

AI and satellites in mining news: dated milestones

Press releases on AI and satellites in mining are easy to find. The ones below come with dates and a named source.

AI in mining reaches past exploration. Rio Tinto’s chief information officer said in 2024 that the company applies these techniques to exploration, digital twins, energy management, mineral extraction and metals processing. Investors who want a wider view of mining names can read our overview of mining investments, our note on Zijin and other mining stocks and, for lithium developers, American lithium companies.

Frequently asked questions

What is AI in mineral exploration?

It is the use of machine-learning models to rank ground for further work, using maps, satellite scenes, geophysics and records of known deposits. The model produces a ranked target list, and field checks and drilling decide whether a target holds ore.

How is AI changing mineral exploration?

Mostly through speed at the desk stage. DARPA and the USGS report assessment runs of two and a half days against a typical two years. Companies such as KoBold also use AI to choose drill targets, though KoBold has yet to bring a mine into production.

Can machine learning find a mineral deposit by itself?

No. It scores ground against patterns from known deposits. The Americas copper study found good results in North America and poorer ones in South America, which shows how much depends on the data and the geology.

What is machine learning mineral exploration used for in Canada?

The Geological Survey of Canada is building national-scale mineral potential models, including carbonatite-hosted rare earth elements and lithium in pegmatites. Companies and consultants apply similar methods to individual districts using provincial geoscience data.

What does AI-enabled exploration mean?

It describes any exploration programme where AI sits in the workflow: ranking targets, processing geophysics, reading old reports or mapping alteration from satellites. Ask which of those jobs is meant and how the output was tested.

What is mineral system modelling?

It treats a deposit as the end of a chain of processes: a source of metal, a pathway, a trap and preservation. Maps of evidence for each link are combined into a target map, and machine learning is one way to do the combining.

What do mining mapping services deliver?

Typical deliverables are geological and alteration maps, structural interpretations and prospectivity maps in GIS formats such as GeoTIFF or shapefile. Ask for inputs, test method and uncertainty to be listed with each map.

Which AI is best for geology?

It depends on the task. Prospectivity models suit districts with many known deposits, geophysics platforms suit drill targeting, language models suit summarising reports, and satellite analysis suits early screening of large or remote ground.

Can satellites find minerals?

Satellites measure reflected light, so they map surface features linked to minerals, such as iron-oxide staining and clay alteration. They do not see ore at depth. Farmonaut states that its reports flag target zones, alteration halos and structures; drilling then tests them.

Is AI in mining the same as crypto mining?

No. Here mining means extracting metals and minerals. AI is used to find deposits and also to plan mines and run plants, as Rio Tinto describes.

Does satellite analysis replace fieldwork?

No. It narrows where to send a field team and what to look for. Field mapping, sampling and drilling remain the proof, and any resource estimate needs a qualified person under NI 43-101 or JORC.