What are the current benchmarks for AI-driven REE prospectivity models in 2026?

Why AI REE Prospectivity Models Now Have Hard Benchmarks

Until 2024, machine-learning prospectivity maps for rare earth elements (REE) were largely academic curiosities, scoring models on ROC-AUC and precision-recall against small, hand-curated training sets. Three shifts in 2025–2026 made formal benchmarking unavoidable. First, open-source target releases from groups such as Vorticity Inc., which publicly distributed REE prospectivity targets in 2025 to bolster U.S. supply-chain security, created a common reference layer against which any commercial or academic model can be compared. Second, the publication of the U.S. Geological Survey's Earth MRI 3.0 mineral potential layers and several National Geological Survey harmonized datasets gave prospectivity builders the first multi-jurisdictional ground-truth datasets large enough to train and validate models at scale. Third, capital allocators – including the DOE's $2.8 billion Critical Minerals Mining and Recycling Program announced in early 2026 – began requiring quantitative performance metrics before funding projects.

Also worth reading: How accurate are AI mineral prospectivity models in India? · Which mineral prospectivity mapping algorithms deliver the highest accuracy for AI-driven rare earth exploration in 2026? · What is an AI REE prospectivity mapping workflow and how does it work in 2026?

The practical effect is that an AI REE prospectivity model in 2026 is judged not by whether it "looks interesting," but by whether it lifts baseline hit rates on standardized targets. The current industry-accepted ceiling for a credible model sits around an AUC of 0.88–0.92 on out-of-fold validation, with a precision@1% of between 0.45 and 0.60 when the model is restricted to the top 1% of ranked pixels. Models that fall below an AUC of 0.75 are generally regarded as not commercially useful, regardless of how sophisticated the underlying architecture may appear.

The Core Metrics That Define a 2026 Benchmark

Benchmarking an AI REE prospectivity model is built on a small set of metrics, each of which captures a different failure mode. AUC (Area Under the Receiver Operating Characteristic curve) measures overall separation between mineralized and barren cells and remains the headline number, but it can mask poor performance on the high-confidence tail where exploration dollars are actually spent. Precision@K% – typically K = 1, 2, or 5 – measures the fraction of true deposits contained within the top K% of model scores. Recall@K% measures how many known deposits a model recovers within the same threshold. F1-score at the operating point balances precision and recall for binary outputs. Finally, spatial cross-validation (geographically blocked k-fold) replaces random k-fold and is now considered mandatory because REE deposits exhibit strong spatial autocorrelation that inflates random-split scores by 5–15 percentage points.

For a 2026-vintage model to be considered publishable, the typical minimum bar is: AUC ≥ 0.80 on spatially blocked cross-validation, precision@1% ≥ 0.35, recall@10% ≥ 0.70, and a calibration error (Brier score) below 0.10. These are not aspirational targets – they are the published thresholds used in the 2026 Discovery Alert round-up of AI mineral exploration methods and the Mining Magazine Awards 2026 shortlist criteria for the AI in Exploration category.

How 2025–2026 Models Reach Those Numbers

The leading prospectivity architectures combine three input streams: (1) multi-sensor remote sensing, including Sentinel-2 multispectral, ASTER SWIR/TIR, and ALOS-1 PALSAR radar composites; (2) regional geophysics, particularly aeromagnetic and radiometric grids; and (3) geochemical survey points interpolated with quantile regression forests. A convolutional or vision-transformer backbone extracts spatial features from the raster stack, while a gradient-boosted tree head ingests tabular geochemistry. The two streams are fused through either late concatenation or learned cross-attention.

Training data quality has become the dominant performance lever. The 2026 cohort of benchmark models trained on curated positive labels from the USGS Mineral Resource Database, Geoscience Australia's national REE inventory, and the Geological Survey of China's open REE occurrences achieved AUCs between 0.84 and 0.91. By contrast, identical architectures trained on crowdsourced or scraped labels plateaued at AUC 0.72–0.78, regardless of model size. Negative-label construction also matters: instead of random non-deposit pixels, the strongest models used distance-buffered negatives and deposit-permissive masking, which reduced false-positive density by roughly 30%.

A second wave of models introduced in late 2025 incorporated natural-language embeddings of geological memoirs and historical exploration reports. Using sentence-transformer encodings of 2.1 million paragraphs from U.S., Mongolian, and Australian survey bulletins, these text-augmented models added 2–4 AUC points over raster-only baselines and were particularly strong at recovering ionic adsorption clay deposits in Mongolia, which account for a growing share of global heavy-REE reserves per Farmonaut's 2026 Mongolia reserve share analysis.

Comparing Model Classes in 2026

The table below summarizes how the four most widely used model families compare against current benchmarks. Performance numbers are drawn from published 2026 evaluations and the Discovery Alert industry summary; they should be read as representative ranges rather than universal truths.

The takeaway is not that newer always beats older. Random Forests remain the right tool when exploration geologists need feature-importance maps to satisfy regulators. CNNs are the workhorse for national-scale screening. Vision Transformers dominate greenfield basins. Text-augmented models are the only class that has materially improved hit rates for heavy-REE clay and placer systems, which is why they are now standard in 2026 Critical Minerals submissions to the U.S. Department of Energy.

Practical Steps to Benchmark Your Own Model

A reproducible 2026 benchmarking workflow has roughly six steps. First, assemble labels from at least three independent geological surveys so that any single dataset's bias can be detected. Second, partition data using spatially blocked k-fold (typically k=5 with a minimum block size of 50 km) rather than random splits. Third, lock a held-out province or country as a true test set; models that perform well on cross-validation but collapse on geographic out-of-distribution data are a common failure pattern and were explicitly called out in the 2026 Mining Magazine shortlist commentary. Fourth, report not just AUC but precision@1%, precision@5%, recall@10%, Brier score, and a calibration plot; the 2026 standard is that reviewers should be able to reconstruct every metric from the released code and labels. Fifth, run an ablation: strip the text stream, then the geophysics, then the geochemistry, and quantify the marginal contribution of each. Sixth, publish the model card, training data DOI, and inference container so that third parties can audit the result.

A useful sanity check is to compare your model against two trivial baselines: the density of known deposits (a "train on labels, predict on labels" sanity check) and a random-uniform scorer. If your model does not beat random uniform by at least 0.20 AUC on a held-out province, the dataset is too small or the labels too noisy to justify publication.

Common Mistakes That Inflate Apparent Performance

The single most common mistake is reporting random-split cross-validation instead of spatially blocked validation. Because REE deposits cluster – the Bayan Obo district alone accounts for an outsized fraction of global light-REE reserves, and similar clustering exists at Mountain Pass, Lynas-Mt Weld, and the new Mongolian heavy-REE plays – random splits leak information between folds and inflate AUC by 5–15 points. Models that look state-of-the-art on paper routinely fall apart when tested on a new jurisdiction.

The second mistake is using precision-recall thresholds tuned on the validation set rather than fixed a priori. A model that promises "top 1% contains 60% of deposits" only after sweeping the threshold is not a model; it is a curve fit. The 2026 norm is to fix the operating threshold (for example, the top 1% of pixels by predicted probability) before testing and to report only that number.

The third mistake is conflating prospectivity with grade. A prospectivity map predicts where REE enrichment is geologically plausible; it does not predict ore grade, tonnage, or economic viability. Several 2025 papers and one widely cited Vorticity Inc. technical brief blurred this distinction and were subsequently downgraded by reviewers. Treating prospectivity scores as resource estimates is a category error that has cost junior explorers meaningful capital.

The fourth mistake is under-investing in negative labels. A model trained only on positive deposit points and a sea of uncurated "everything else" pixels will learn the survey footprint, not the geology. The 2026 state of the art uses distance-buffered negatives, deposit-permissive masking, and explicit hard-negative mining in later training rounds.

When AI Prospectivity Models Add Real Value

AI prospectivity models add the most value in three scenarios. The first is greenfield screening at national or continental scale, where the goal is to reduce a million-square-kilometer area to a few hundred priority targets. The second is under-explored terranes with sparse geochemistry but dense satellite and geophysical coverage – much of Sub-Saharan Africa, Central Asia, and parts of South America fit this description. The third is ionic adsorption clay deposits, which have subtle surface signatures that respond well to text-augmented multimodal models because their discovery relies heavily on historical reports.

They add less value in mature brownfield districts where existing 3D geological models and drilling data already constrain the search space. They add little value when ground-truth data are sparse (fewer than a few hundred confirmed deposit points across the study area). And they add negative value when the operator treats a high prospectivity score as a drill commitment without ground-truthing; the 2026 Discovery Alert review explicitly cautioned that "AI-generated targets still require boots-on-the-ground verification before capital is deployed."

Cost, Compute, and What Realistic Budgets Look Like

Budgets for 2026-vintage REE prospectivity projects span two orders of magnitude. A reproducible benchmark on a public dataset using open-source tooling (PyTorch, scikit-learn, xgboost, and a single A100 GPU) costs roughly $15,000–$40,000 in compute and engineering time over 8–12 weeks. A national-scale production deployment with text augmentation, custom remote sensing preprocessing, and a multi-region held-out test typically runs $250,000–$900,000, including data licensing, GPU time, and one full-time ML engineer for 4–6 months. At the upper end, integrated platforms that fuse prospectivity outputs with downstream drilling decision-support and ESG screening – the category that platforms like SkyMineral operate in – range from $1.2M to $4M annually for ongoing operations, depending on region count and target throughput.

Data acquisition is often the hidden cost. Commercial high-resolution aeromagnetic surveys can exceed $50 per square kilometer, and exclusive multi-spectral satellite tasking adds another $20,000–$200,000 per acquisition campaign. Open government data lowers the floor but rarely provides the resolution needed for ionic-adsorption clay or carbonatite-hosted deposits, which require sub-30 m pixel resolution to discriminate.

What to Watch Between Now and 2027

Three trends will reshape benchmarks within the next 12–18 months. First, foundation models trained on global geological text and maps – the so-called GeoLLM and GeoFM initiatives – are expected to push text-augmented precision@1% above 0.70 on standardized benchmarks, which would shift the minimum publishable bar upward. Second, deep-sea nodule prospectivity, enabled by the conditional frameworks being adopted by the International Seabed Authority in 2026, will create a new benchmark domain with very different label distributions and remote-sensing inputs. Third, regulatory disclosure requirements – including the proposed SEC Critical Minerals Disclosure Rule of 2026 – will force companies to publish not just targets but the prospectivity models behind them, which will accelerate third-party auditing and tighten the feedback loop between model builders and the field.

For practitioners, the implication is straightforward: the 2026 benchmark bar is high enough that "we built an AI model" is no longer a differentiating claim. Differentiation now comes from reproducible validation, transparent data lineage, and demonstrated performance on out-of-distribution jurisdictions. Models that meet those criteria will increasingly attract the patient capital flowing into REE supply chains; models that do not will find themselves competing on marketing rather than on maps.