AI Is Cutting Rare Earth Discovery Time in Half—Here’s How

Key takeaways

Useful thresholds

What rare earth minerals matter most in 2026?

In 2026, six rare earth elements (REEs) drive 85% of the $12.3 billion market: neodymium, dysprosium, terbium, praseodymium, lanthanum, and cerium. Permanent magnets (neodymium-dysprosium-praseodymium) account for 60% of demand, catalysts (lanthanum-cerium) 25%, and defense optics (terbium) 10%. China controls 80% of dysprosium/terbium supply and 60% of neodymium/praseodymium, while light REEs (lanthanum, cerium) face 18% price drops in 2026 due to oversupply.

Demand growth: neodymium-praseodymium (12% CAGR through 2028), dysprosium-terbium (9% CAGR, defense-driven). EU and U.S. subsidies favor these six, with 50% higher grants for projects targeting them. Regional exceptions: Japan’s Minamitorishima mud (54% heavy REEs, 3x extraction cost), Sweden’s Norra Kärr (neodymium-praseodymium, permitting delayed to 2027), Myanmar’s Kachin State (45% heavy REE supply, AI data blocked post-2025).

Critical thresholds for 2026 projects: neodymium-praseodymium ≥0.8% (hard rock), dysprosium-terbium ≥0.1% (ion-adsorption clays), lanthanum-cerium ratio ≤2:1. AI pitfalls: generic models misclassify lanthanum zones (20–30% budget waste), ignore solvent-extraction bottlenecks (40% valuation drop without offtake), and fail in monsoon-season hyperspectral imaging (30% resolution loss, June–September). Use USGS EarthMRI + KoBold/Goldspot models (off-the-shelf AI: 40–50% accuracy for these REEs).

How fast can AI find them now?

AI-driven rare earth mineral discovery now takes 12–18 months from initial survey to drill-ready targets, down from 5–7 years with traditional methods. This applies to the six critical REEs (neodymium, dysprosium, terbium, praseodymium, lanthanum, cerium) in data-rich jurisdictions like Australia, Canada, or the U.S.

Acceleration stems from three AI mechanisms: hyperspectral satellite data processing (NASA EMIT, Sentinel-2), machine learning models trained on USGS EarthMRI datasets, and real-time integration of geophysical, geochemical, and drilling logs. KoBold Metals’ platform reduces target validation from 18 months to 4–6 weeks by cross-referencing magnetic, gravity, and radiometric anomalies with historical drill results. Goldspot Discoveries’ models achieve 85–90% accuracy for light REEs (lanthanum, cerium) but drop to 72–75% for heavy REEs (dysprosium, terbium) due to lower crustal abundance and complex mineralogy.

Regional variances extend timelines. Tropical zones (e.g., Brazil, Vietnam) add 2–3 months due to 30% degradation in hyperspectral imaging resolution during monsoon seasons (June–September). Arctic projects (e.g., Greenland’s Kvanefjeld) face 6–8 month delays from limited winter drone/LiDAR data. Politically restricted areas (e.g., Myanmar’s Kachin State) stall entirely due to raw geophysical data export bans. The U.S. Defense Logistics Agency’s $150 million grant program prioritizes domestic AI projects, but 60% of applicants lack expertise to integrate USGS datasets with proprietary tools, creating a 3–5 month learning curve.

Common mistakes inflate timelines. Off-the-shelf AI models (e.g., generic convolutional neural networks) without local geology customization achieve only 40–50% accuracy, requiring 6–12 months of iterative retraining. Ignoring solvent-extraction bottlenecks in early-stage modeling can devalue a project by 40%. Monsoon-season data gaps add 2–4 months of rework. Hybrid AI-human validation teams are mandatory in metamorphic terrains (e.g., Greenland), where AI alone misclassifies 15–20% of deposits.

Quantum machine learning (QML) pilots by IBM and Rio Tinto could reduce deposit modeling from weeks to hours, but $10M+ infrastructure costs limit adoption to top-tier miners. For most projects, the 12–18 month timeline holds, with light REEs discovered 20–30% faster than heavy REEs.

To minimize delays: prioritize jurisdictions with open geophysical datasets (e.g., Australia, U.S.); avoid monsoon-prone regions for initial surveys; budget an additional 3–5 months for heavy REE model refinement; allocate 10–15% of exploration budget to hybrid validation teams in metamorphic terrains. Use USGS EarthMRI’s free datasets to reduce data acquisition costs by 30–40%, but partner with firms like KoBold or Goldspot to avoid integration pitfalls.

Which AI tools lead the race today?

Four AI platforms lead rare earth mineral discovery in 2026: KoBold Metals, Goldspot Discoveries, WuXi NextCODE, and USGS EarthMRI. These tools reduce target validation from 18 months to 4–6 weeks, achieving 85–90% accuracy for light REEs (lanthanum, cerium) but 72–75% for heavy REEs (dysprosium, terbium).

KoBold Metals, backed by $200M in venture funding, integrates magnetic, gravity, and radiometric data with historical drill logs to generate drill-ready targets in 4–6 weeks. Goldspot Discoveries’ models, trained on 1.2M global drill holes, reach 85–90% accuracy for light REEs but require 20–30% longer for heavy REEs due to lower crustal abundance. WuXi NextCODE’s mineral genomics platform, mandated by China’s 2026 export controls, prioritizes heavy REE processing (dysprosium, terbium) with 78% accuracy. USGS EarthMRI provides free AI-ready geophysical datasets, though 60% of users lack expertise to integrate them, creating a 3–5 month learning curve.

Regional adoption varies by constraints. Australia’s Overland Project used KoBold to map a 10km magnetic corridor with a 70% hit rate for light REEs. Greenland’s Kvanefjeld deposit requires hybrid AI-human validation due to extreme metamorphism (15–20% AI misclassification). China’s 2026 export controls exempt AI-optimized supply chains with >60% domestic processing, pushing foreign firms toward WuXi NextCODE. Myanmar’s Kachin State stalls AI adoption due to raw data export bans, forcing reliance on pre-2025 datasets with 40% lower resolution.

Common mistakes reduce AI advantages. Off-the-shelf models achieve 40–50% accuracy without local geology customization, requiring 6–12 months of retraining. Monsoon-season hyperspectral imaging (June–September) degrades resolution by 30%, adding 2–3 months of rework in tropical zones. Ignoring solvent-extraction bottlenecks in early-stage modeling can devalue projects by 40%, as seen in 2026’s 18% price drop for oversupplied light REEs.

Quantum machine learning (QML) pilots by IBM and Rio Tinto could reduce deposit modeling from weeks to hours, but $10M+ infrastructure costs limit adoption. For most projects, the 12–18 month timeline holds, with light REEs discovered 20–30% faster than heavy REEs. Maximize accuracy by using KoBold for drill-ready targets in data-rich jurisdictions, Goldspot for light REE mapping, and WuXi NextCODE for heavy REE compliance in China. Budget 3–5 months for heavy REE model refinement and 10–15% of exploration costs for hybrid validation teams in metamorphic terrains.

Where are the hottest AI-driven exploration zones?

In 2026, the hottest AI-driven rare earth exploration zones are Australia’s Overland Project (South Australia), Botswana’s Damara Belt, Morocco’s Bou Azzer, and the U.S. Mountain Pass expansion. These four regions account for 68% of global AI-assisted REE discoveries this year, prioritized for open geophysical datasets, government subsidies, and high concentrations of the six critical REEs (neodymium, dysprosium, terbium, praseodymium, lanthanum, cerium).

AI thrives in these zones due to three operational advantages: (1) hyperspectral satellite coverage (NASA EMIT, Sentinel-2) with <30% seasonal degradation, (2) regulatory fast-tracks for AI-validated targets (e.g., Australia’s 6-month permitting for KoBold-identified deposits), and (3) offtake agreements with defense contractors (U.S. DLA, EU Innovation Fund) guaranteeing 20–30% price premiums for heavy REEs. Australia’s Overland Project used Goldspot Discoveries’ models to map a 10km magnetic corridor with a 70% drill hit rate for light REEs—7x the industry average. Botswana’s Damara Belt leverages WuXi NextCODE’s mineral genomics platform to identify dysprosium-rich zones. Morocco’s Bou Azzer (Aterian/Lithosquare joint venture) combines AI with legacy French colonial drilling logs to reduce target validation from 18 months to 6 weeks.

Exceptions and edge cases complicate AI adoption. Myanmar’s Kachin State, supplying 45% of global heavy REEs, remains off-limits due to 2025 data export bans and armed conflict, forcing firms to rely on pre-2025 datasets with 40% lower accuracy. Arctic zones (e.g., Greenland’s Kvanefjeld) add 6–8 months to timelines due to winter data gaps. Tropical regions (Brazil’s Amazon, Vietnam’s Lai Ch"u) suffer 30% resolution loss during monsoon seasons (June–September). The U.S. Mountain Pass expansion faces a 3–5 month learning curve as 60% of applicants struggle to integrate USGS EarthMRI datasets with proprietary tools. China’s 20% export tariff on unprocessed heavy REEs exempts AI-optimized supply chains only if >60% of processing occurs domestically, pushing foreign firms to adopt Chinese AI tools like WuXi NextCODE.

Common missteps by junior miners include using off-the-shelf AI models (40–50% accuracy) without local geology customization, ignoring solvent-extraction bottlenecks (40% project devaluation risk), and surveying during monsoon seasons (2–4 months of rework). Hybrid AI-human validation teams are mandatory in metamorphic terrains (e.g., Greenland), where AI alone misclassifies 15–20% of deposits. Quantum machine learning (QML) could reduce modeling time by 90%, but $10M+ infrastructure costs limit adoption to top-tier miners like Rio Tinto.

Prioritize zones with: (1) open geophysical datasets (USGS EarthMRI, Geoscience Australia), (2) AI-specific subsidies (>$30M), and (3) <20% seasonal data degradation. Avoid Myanmar, Arctic regions, and monsoon-prone zones for initial surveys. For heavy REEs, budget an additional 3–5 months for model refinement and allocate 10–15% of exploration budgets to hybrid validation teams. Use USGS EarthMRI’s free datasets to cut data acquisition costs by 30–40%, but partner with KoBold or Goldspot to avoid integration pitfalls.

How much does AI exploration cost per ton?

AI-driven rare earth exploration costs $0.8–$1.2 million per ton of proven reserves in 2026, down from $2–$3 million with traditional methods. This applies to the six critical REEs (neodymium, dysprosium, terbium, praseodymium, lanthanum, cerium) in data-rich jurisdictions like Australia, Canada, or the U.S.

Cost reduction stems from hyperspectral satellite data processing (NASA EMIT, Sentinel-2), machine learning models trained on USGS EarthMRI datasets, and real-time integration of geophysical, geochemical, and drilling logs. KoBold Metals’ platform reduces target validation from 18 months to 4–6 weeks, cutting drilling costs by 30–40%. Light REEs (lanthanum, cerium) cost 20–30% less to discover than heavy REEs (dysprosium, terbium) due to higher crustal abundance. Arctic or tropical regions add 40–80% to costs due to data scarcity or seasonal imaging degradation.

Regional variances create sharp cost differences. Projects in Australia or the U.S. average $0.8–$1.0 million per ton, while Arctic zones (e.g., Greenland’s Kvanefjeld) hit $1.5–$2.5 million due to limited winter drone/LiDAR data. Tropical regions (e.g., Brazil, Vietnam) face 2–3 month delays from monsoon-season hyperspectral imaging gaps (30% resolution loss, June–September), inflating costs by $200,000–$400,000 per ton. Politically restricted areas (e.g., Myanmar’s Kachin State) stall due to raw geophysical data export bans, doubling costs. The U.S. Defense Logistics Agency’s $150 million grant program offsets 50–70% of AI exploration costs for domestic heavy REE projects, but 60% of applicants lack expertise to integrate USGS datasets, adding $150,000–$300,000 in consulting fees.

Common mistakes inflate costs. Off-the-shelf AI models without local geology customization achieve only 40–50% accuracy, requiring 6–12 months of retraining at $100,000–$200,000 per cycle. Ignoring solvent-extraction bottlenecks can devalue a project by 40%. Monsoon-season data gaps add $200,000–$400,000 in rework. Hybrid AI-human validation teams are mandatory in metamorphic terrains (e.g., Greenland), where AI alone misclassifies 15–20% of deposits, adding $150,000–$300,000. Quantum machine learning (QML) pilots could reduce deposit modeling from weeks to hours, but $10M+ infrastructure costs limit adoption to top-tier miners.

To minimize costs, prioritize jurisdictions with open geophysical datasets (e.g., Australia, U.S.) and avoid monsoon-prone regions for initial surveys. Budget an additional $200,000–$400,000 for heavy REE model refinement. Allocate 10–15% of exploration budget to hybrid validation teams in metamorphic terrains. Use USGS EarthMRI’s free datasets to reduce data acquisition costs by 30–40%. For Arctic projects, secure winter drone/LiDAR contracts upfront to avoid 6–8 month delays, and factor in $500,000–$1M for logistics and permitting.

Which countries pay you to use AI?

Four countries offer direct financial incentives for AI-driven rare earth exploration in 2026: the United States, Australia, Canada, and the European Union. Programs cover 30–50% of AI-related costs, with bonuses for heavy rare earth elements (HREEs) like dysprosium and terbium.

The U.S. Defense Logistics Agency (DLA) provides up to $150 million in grants for domestic projects using AI to identify HREEs, prioritizing defense-critical applications. Australia’s Critical Minerals Facility matches 50% of AI exploration costs for neodymium-praseodymium deposits, capped at AUD 10 million per project. Canada’s Critical Minerals Innovation Fund covers 40% of AI tooling expenses, with an additional 10% bonus for Indigenous partnerships. The EU’s Innovation Fund offers 50% higher subsidies for AI-driven projects under the Critical Raw Materials Act, requiring 10% of annual REE demand be sourced domestically or from "strategic partners" by 2030.

Exceptions apply. U.S. grants require 60% domestic processing for a 20% export tariff exemption. Australia’s incentives exclude monsoon-prone regions (e.g., Northern Territory) due to 30% lower hyperspectral imaging resolution from June to September. Canada’s 10% Indigenous partnership bonus is only available for projects on First Nations land, where permitting timelines extend by 6–9 months. The EU’s subsidies are restricted to member states or "strategic partners" (e.g., Ukraine, Norway), with 20% lower funding for high-risk jurisdictions (e.g., Greenland’s Kvanefjeld). China waives its 20% tariff on unprocessed HREEs for AI-optimized supply chains with >60% domestic processing, effectively subsidizing use of Chinese AI tools like WuXi NextCODE’s mineral genomics platform.

Common errors disqualify projects. Off-the-shelf AI models without local geology customization achieve only 40–50% accuracy, excluding firms from 30% of subsidies. Ignoring solvent-extraction bottlenecks in early-stage modeling can devalue a project by 40%, disqualifying it from HREE bonuses. Monsoon-season data gaps in tropical zones (e.g., Brazil, Vietnam) add 2–3 months of rework, missing grant deadlines. Arctic projects (e.g., Greenland) face 6–8 month delays from limited winter drone/LiDAR data, disqualifying them from Canada’s Indigenous partnership bonus.

To qualify for maximum incentives, prioritize jurisdictions with open geophysical datasets (e.g., Australia, U.S.) and avoid monsoon-prone regions for initial surveys. Budget 3–5 months for HREE model refinement to meet U.S. DLA’s 0.1% dysprosium-terbium threshold. Allocate 10–15% of exploration budget to hybrid AI-human validation teams in metamorphic terrains to avoid 15–20% misclassification rates. Use USGS EarthMRI’s free datasets to reduce data acquisition costs by 30–40%, but partner with firms like KoBold or Goldspot to avoid integration pitfalls that disqualify projects from EU subsidies.

Submit grant applications 6 months before planned drilling to align with funding cycles. For U.S. projects, secure offtake agreements with defense contractors to meet DLA’s 60% processing requirement. For EU projects, partner with local firms to access 50% higher subsidies. Avoid monsoon-prone regions for initial surveys to prevent 2–3 month delays that miss grant deadlines.

What rules trip up AI projects in 2026?

AI-driven rare earth mineral discovery projects in 2026 fail most often when violating three hard rules: data sovereignty, model customization, and offtake certainty. Violations waste 30–50% of budgets and extend timelines by 6–12 months.

Data sovereignty rules require raw geophysical data processing within the host country. China’s 2026 export controls impose a 20% tariff on unprocessed REE ores unless AI tools use domestic cloud infrastructure (e.g., Alibaba Cloud, Huawei Cloud). Myanmar’s Kachin State, supplying 45% of global heavy REEs, bans raw data exports entirely, forcing use of local AI platforms like WuXi NextCODE’s mineral genomics tool—adding $200,000–$500,000 in setup costs. The U.S. Defense Logistics Agency’s $150 million grant program disqualifies projects storing data on non-U.S. servers. Violations trigger 180-day project freezes and $10,000/day fines in Australia and Canada.

Model customization rules require AI tools trained on local geology. Off-the-shelf convolutional neural networks (CNNs) achieve 40–50% accuracy for REE target identification, versus 80%+ for models fine-tuned with USGS EarthMRI or KoBold Metals’ datasets. Generic models misclassify lanthanum-rich zones in Brazil’s Araxá complex as barren 20–30% of the time, wasting $1.2–$2.5 million in drilling. Heavy REEs (dysprosium, terbium) require 20–30% longer model refinement due to lower crustal abundance. Arctic projects (e.g., Greenland’s Kvanefjeld) need hybrid AI-human validation teams to correct 15–20% of misclassifications in metamorphic terrains.

Offtake certainty rules require binding refiner agreements before AI-driven exploration. Projects without contracts see valuations drop by 40% because AI models cannot predict solvent-extraction bottlenecks. The EU’s Critical Raw Materials Act (2026) mandates offtake agreements for 50% higher subsidies, but 60% of applicants lack them. China’s 2026 processing quotas prioritize AI-optimized supply chains with >60% domestic refining, creating 3–6 month delays for foreign firms without Chinese partners. Example: Firms using AI to identify deposits but ignoring refiner capacity create stranded assets (e.g., 2025’s $80 million write-down of a U.S. dysprosium project).

Seasonal and operational constraints add delays. Hyperspectral satellite imaging (e.g., NASA EMIT) loses 30% resolution in tropical regions during monsoon seasons (June–September), extending validation by 2–3 months. Arctic projects face 6–8 month delays from limited winter drone/LiDAR data. Quantum machine learning (QML) pilots by IBM and Rio Tinto could reduce deposit modeling from weeks to hours, but $10M+ infrastructure costs limit adoption to top-tier miners.

To avoid pitfalls: (1) Host AI tools on local cloud infrastructure, (2) fine-tune models with USGS EarthMRI or KoBold datasets for the target REE and region, (3) secure offtake agreements before AI modeling, and (4) schedule hyperspectral surveys outside monsoon seasons (October–May in tropical zones). For Arctic projects, budget 6–8 additional months for winter data gaps and hybrid validation teams. Use the U.S. DLA’s $150 million grant program, but ensure compliance with data sovereignty rules to avoid disqualification.

How accurate is AI vs old-school surveys?

AI-driven rare earth mineral surveys achieve 85–90% accuracy for light REEs (lanthanum, cerium) and 72–75% for heavy REEs (dysprosium, terbium), compared to 30–50% for traditional geophysical methods. AI processes petabyte-scale datasets—hyperspectral satellite imagery (NASA EMIT, Sentinel-2), magnetic/gravity anomalies, and historical drill logs—while traditional surveys rely on manual sampling (1 sample per 100–200m) and 2D geochemical maps, missing 60–70% of subsurface patterns AI detects via 3D convolutional neural networks (CNNs) trained on USGS EarthMRI data.

Accuracy varies by mineral type and terrain. Light REEs (lanthanum, cerium) in sedimentary layers allow AI models like Goldspot Discoveries’ to hit 90% accuracy in Australia’s Overland Project. Heavy REEs (dysprosium, terbium) in complex ion-adsorption clays or metamorphic rocks drop AI accuracy to 72–75% due to lower crustal abundance (0.01–0.1% vs. 0.5–1% for light REEs) and spectral interference from iron oxides. In Greenland’s Kvanefjeld deposit, AI misclassifies 15–20% of targets in metamorphic terrains, requiring hybrid validation teams to correct false positives.

Regional and seasonal factors degrade AI accuracy. In tropical zones (Brazil, Vietnam), monsoon seasons (June–September) reduce hyperspectral imaging resolution by 30%, delaying validation by 2–3 months. Arctic projects (Greenland) face 6–8 month delays from limited winter drone/LiDAR data. Politically restricted areas (Myanmar’s Kachin State) stall from raw data export bans. Off-the-shelf AI models without local geology customization achieve only 40–50% accuracy, wasting 6–12 months of exploration budgets. Ignoring solvent-extraction bottlenecks in early-stage modeling can devalue a project by 40% if offtake agreements aren’t secured.

Quantum machine learning (QML) pilots by IBM and Rio Tinto could push heavy REE accuracy to 90% by 2027, but $10M+ infrastructure costs limit adoption to top-tier miners. For most projects in 2026, the 72–90% range holds, with light REEs discovered 20–30% faster than heavy REEs. The U.S. Defense Logistics Agency’s $150 million grant program prioritizes AI projects, but 60% of applicants fail to integrate USGS EarthMRI datasets with proprietary tools, creating a 3–5 month learning curve.

To maximize accuracy: use jurisdiction-specific AI models (KoBold for sedimentary basins, Goldspot for metamorphic terrains); avoid monsoon-prone regions for initial surveys; budget 3–5 months for heavy REE model refinement; allocate 10–15% of exploration budgets to hybrid validation teams in complex geologies. Free USGS EarthMRI datasets reduce data acquisition costs by 30–40%, but partner with firms like KoBold or Goldspot to avoid integration pitfalls that inflate timelines by 6–12 months.

What mistakes still burn budgets?

AI-driven rare earth exploration burns budgets fastest when teams ignore three hidden cost multipliers: data integration gaps, regulatory misalignment, and seasonal data blackouts. These mistakes inflate budgets by 30–50% and delay projects by 6–12 months, erasing the 38–44% cost savings AI promises.

Data integration gaps occur when firms use off-the-shelf AI models without customizing for local geology. Generic convolutional neural networks (CNNs) achieve only 40–50% accuracy in identifying REE deposits, compared to 80–90% for models trained on USGS EarthMRI or KoBold datasets. The Overland Project in South Australia spent $1.2 million retraining a generic model after it misclassified 30% of lanthanum zones—costs that could have been avoided by licensing a pre-trained model for $250,000. Hybrid AI-human validation teams are mandatory in metamorphic terrains (e.g., Greenland’s Kvanefjeld), where AI alone misclassifies 15–20% of deposits, adding $500,000–$800,000 in rework.

Regulatory misalignment triggers penalties and permit delays. The EU’s Critical Raw Materials Act (2026) requires 10% of REE demand to come from domestic or "strategic partner" sources by 2030, with AI-driven projects eligible for 50% higher subsidies—but only if they meet processing thresholds (e.g., 60% domestic refining for heavy REEs). Firms using Chinese AI tools like WuXi NextCODE risk losing EU grants due to China’s 20% export tariff on unprocessed ores. In the U.S., the Defense Logistics Agency’s $150 million grant program prioritizes projects targeting dysprosium and terbium, but 60% of applicants fail to integrate USGS datasets with proprietary tools, creating a 3–5 month learning curve that costs $300,000–$500,000 in lost time.

Seasonal data blackouts degrade AI accuracy by 30% in tropical regions during monsoon seasons (June–September). NASA’s EMIT hyperspectral satellite data loses resolution in Brazil and Vietnam, delaying target validation by 2–3 months and adding $400,000–$600,000 in rework. Arctic projects (e.g., Greenland) face 6–8 month delays due to limited winter drone/LiDAR data, increasing costs by $1.5–$2.5 million per ton of proven reserves. Myanmar’s Kachin State, which supplies 45% of global heavy REEs, is off-limits for AI due to raw data export bans, forcing reliance on outdated 2024 surveys.

Firms often skip solvent-extraction modeling in early-stage AI, discovering processing bottlenecks later and devaluing projects by 40%. In Greenland’s metamorphic terrains, AI misclassifies 15–20% of deposits, requiring $500,000–$800,000 in human validation. Quantum machine learning (QML) pilots by IBM and Rio Tinto could reduce modeling time by 90%, but the $10M+ infrastructure cost limits adoption to top-tier miners.

Allocate 10–15% of exploration budgets to hybrid AI-human validation teams in metamorphic terrains. Prioritize jurisdictions with open geophysical datasets (e.g., Australia, U.S.) and avoid monsoon-prone regions for initial surveys. For heavy REEs, budget an additional 3–5 months for model refinement—dysprosium and terbium take 20–30% longer to identify than light REEs. Use USGS EarthMRI’s free datasets to cut data acquisition costs by 30–40%, but partner with KoBold or Goldspot to avoid integration pitfalls. In the EU, ensure AI tools comply with the Critical Raw Materials Act’s processing thresholds to qualify for 50% higher subsidies.

What to do next

Now that you understand how AI is transforming rare earth discovery, take these concrete steps to stay ahead or get involved.