How is AI drill planning changing rare earth exploration in 2026?

AI drill planning has moved from an experimental curiosity to a working part of the rare earth and critical minerals toolkit, and by August 2026 it is being deployed on real drill programs rather than discussed in whitepapers. The clearest recent example is Canamera's decision to deploy ExploreTech's Stanford-born AI drill planning platform at Schryburt Lake ahead of its maiden drill program, announced via TMX Newsfile. That single deployment captures the shift: junior explorers, who historically planned holes with paper maps, cross-sections, and the intuition of one senior geologist, are now handing targeting decisions to machine learning models trained on geophysics, geochemistry, structural data, and historical drilling results.

What AI Drill Planning Actually Does

Also worth reading: How does AI drill target optimization work in mineral exploration, and is it worth using before a maiden drill program? · What is the future of AI mineral exploration, and how will it change how we find rare earths and critical minerals? · Can predictive models replace traditional geological fieldwork and drill planning?

At its core, AI drill planning answers three questions that every exploration manager faces: where should we drill next, in what order, and when should we stop? Traditional workflows answer these questions through manual interpretation of magnetic, gravity, radiometric, and electromagnetic surveys combined with geochemical assays from prior holes. A senior geologist might spend weeks building a 3D model and then argue about collar locations in a conference room. An AI platform ingests the same datasets plus every historical hole within the region, learns the statistical signatures associated with mineralization, and outputs ranked targets with quantified uncertainty.

The practical difference shows up in hit rates and cost per discovery ounce or kilogram. Industry analyses over the past several years have suggested that AI-assisted targeting can reduce the number of holes required to reach a discovery by 30 to 50 percent in favorable settings, because the model prioritizes high-probability zones instead of drilling on a grid pattern for coverage's sake. For a rare earth project, where a maiden program can run $3 million to $15 million depending on location and rig availability, cutting even ten holes from a forty-hole program saves real money. The caveat, which vendors rarely emphasize, is that these statistics come from case studies selected after the fact; nobody publishes the AI campaigns that failed.

The technology stack matters too. Platforms differ in whether they use supervised learning (trained on known deposits), unsupervised anomaly detection (flagging anything statistically unusual), or physics-informed models that respect geological constraints like fault geometry. ExploreTech, spun out of Stanford research, sits in the physics-informed camp, which tends to produce recommendations geologists find defensible. Paris-based Lithosquare raised €22 million to scale what it calls Geology AI for transition-critical mineral discovery, signaling that venture capital now treats this as an infrastructure play rather than a niche software market.

Why Rare Earths Are a Special Case

Rare earth elements present targeting problems that generic mineral AI handles poorly, and this is worth understanding before anyone signs a contract. Unlike gold, where a single element defines the target, rare earth deposits are evaluated on a basket: neodymium and praseodymium carry most of the magnet-market value, while cerium and lanthanum often dominate tonnage but add little revenue. An AI model trained only on total REO (rare earth oxide) grades will happily recommend drilling zones full of low-value cerium. The best platforms therefore optimize for NdPr oxide equivalent or magnet-relevant value density, not raw grade.

Deposit type adds another layer. Carbonatites (like Mountain Pass in California), ion-adsorption clays (historically dominant in southern China and now targeted across Southeast Asia and South America), monazite-bearing heavy mineral sands, and alkaline igneous complexes all have distinct geophysical and geochemical fingerprints. Heavy mineral sands are the easiest case for AI because they are shallow, laterally extensive, and well-sampled — which is why Rare Earths Americas could confidently intercept similar sands more than 50 kilometers from its Shiloh deposit and treat it as a district-scale extension rather than a gamble. Hard-rock carbonatite targeting is harder, requiring the model to integrate deep geophysics with limited outcrop data.

There is also a geopolitical accelerant. China still controls the large majority of global rare earth refining and magnet production, and the United States, Japan, and Europe are spending heavily to change that. Japan has stepped up its deep-sea mining plan specifically to cut rare-earth dependence on Chinese supply chains, per Bloomberg reporting. Greenland has become a focal point, with investors including figures like Jeff Bezos, Bill Gates, and Sam Altman following US Arctic ambitions, as Forbes and TradingView coverage of tickers such as CRML, UUUU, USAR, ALOY, and GLND illustrates. When governments subsidize exploration and processing — the Brook Mine in Wyoming, for instance, plans a rare earth and critical minerals processing plant designed by Fluor Corporation — the economics of aggressive, AI-accelerated drilling improve dramatically, because time-to-resource becomes a strategic metric rather than just a budget line.

Comparison: AI Platforms vs. Traditional Targeting

Neither column wins outright. On a grassroots property with two soil samples and an old airborne survey, an AI model has almost nothing to learn from, and a good field geologist will outperform it. Where AI earns its fee is on mature properties with decades of accumulated data — exactly the situation at Schryburt Lake, where Canamera had enough prior work to make machine learning worthwhile before committing rig money to a maiden program.

Practical Steps to Deploy AI Drill Planning

The first step is a data audit, and it is unglamorous. Most juniors discover their historical drill logs exist as scanned PDFs, their assays live in inconsistent CSV formats, and their geophysics is in proprietary vendor formats. Before any model runs, someone must standardize coordinate systems, flag duplicate records, and digitize legacy logs. Budget four to eight weeks and $20,000 to $100,000 for this on a mid-sized project; skipping it produces garbage-in-garbage-out predictions that quietly poison the whole campaign.

Second, define the objective function explicitly. If the goal is NdPr-rich carbonatite, say so in the model specification and weight training labels accordingly. If the goal is total REO tonnage for a potential government-backed supply deal, that is a different optimization. Vague objectives produce vague targets, and vague targets produce arguments between the AI vendor and the chief geologist that stall the program.

Third, run a retrospective validation. Take the existing dataset, hide the results of the last twenty drilled holes, let the model predict them, and measure how many it would have hit. This back-testing step costs little and reveals whether the platform actually works on your geology or merely works on the vendor's demo dataset. Reputable providers welcome this test; if a vendor resists back-testing, treat that as a red flag worth walking away from.

Fourth, keep humans in the loop on collar placement. The output should be ranked candidate zones, not final coordinates. Structural geologists routinely catch errors — a modeled fault that contradicts field mapping, a magnetic anomaly that is clearly cultural interference from an old fence line. The workflow that works is AI proposes, geology disposes.

Fifth, instrument the drill program itself. Log actual versus predicted results hole by hole so the model retrains on fresh data mid-campaign. Programs that retrain weekly consistently outperform those that lock the model at kickoff, because early holes reveal systematic biases in the original interpretation.

Common Mistakes and Failure Modes

The most expensive mistake is treating AI output as ground truth. Several well-funded startups have burned investor capital by drilling purely model-ranked targets without field verification, then discovering the anomalies were artifacts of incomplete geophysical coverage rather than mineralization. Machine learning finds statistical patterns; it does not know that a lake sediments anomaly reflects glacial transport from ten kilometers away.

A second mistake is underweighting metallurgy. A discovery is not a mine, and rare earths are notorious for metallurgical complexity — mineralogy determines whether a deposit is economically recoverable far more than headline grade does. AI can rank drill targets beautifully and still deliver a resource that no flowsheet processes profitably. Teams should pair targeting AI with automated mineralogy (QEMSCAN-style analysis) from the first campaign, not after the resource estimate.

Third, teams misjudge jurisdictional and ESG friction. Greenland's minerals story attracts billionaires and headlines, but permitting timelines there remain long and politically volatile, as CNBC's coverage of tech investors assessing mining amid US takeover talk makes clear. Europe's push to develop what CNBC called blockbuster rare earths discoveries faces similar delays. An AI-optimized drill schedule that ignores a two-year permitting queue optimizes the wrong bottleneck entirely.

Fourth, cost expectations get distorted. Platform licensing is rarely the biggest expense; data remediation, additional geophysics to feed the model, and extended field seasons to chase AI-flagged follow-up targets frequently exceed the software bill. Model the total program cost, not the subscription fee.

When to Act and What It Costs

Timing favors adoption right now for three reasons. First, government money is flowing: US, Japanese, and European programs are subsidizing critical minerals supply chains, meaning partially de-risked economics for domestic exploration. Second, the vendor market has consolidated enough that credible options exist — ExploreTech at the physics-informed end, Lithosquare with fresh €22 million in funding, and several in-house efforts at majors — but pricing has not yet hardened into enterprise-software territory everywhere. Third, rig availability remains tight in hot districts, so programs that drill fewer, better-placed holes get rigs sooner.

On cost, expect a spectrum. A pilot study on existing data typically runs $50,000 to $150,000. A full deployment supporting a 20-to-40-hole program, including data engineering and ongoing model updates, commonly lands between $250,000 and $750,000 — meaningful for a junior, trivial against a $10 million drill budget if it removes even five wasted holes. Majors increasingly build internal capability instead, hiring data science teams directly, which trades higher fixed cost for control and confidentiality.

For investors evaluating companies claiming AI-driven discovery, ask three questions: was the model back-tested against blind holes, what objective function was optimized, and how many holes did the AI actually change compared to the geologist's original plan? Companies with honest answers to all three deserve attention. Companies that use AI as a marketing word without methodology disclosure do not. The technology is real, the deployments are real, and the savings are measurable — but the gap between genuine capability and promotional language is wide, and diligence is still the investor's job.