# What metrics matter most when using machine learning for rare earth mineral targeting?

skymineral.com · August 25, 2026

> Machine learning rare earth targeting metrics are the quantitative measures used to evaluate how well an algorithm predicts where rare earth element...

Machine learning rare earth targeting metrics are the quantitative measures used to evaluate how well an algorithm predicts where rare earth element (REE) deposits are likely to occur. The core set includes prospectivity mapping accuracy scores such as AUC-ROC (typically 0.80–0.95 for well-constrained models), precision and recall at specific prediction thresholds, prediction-area (P-A) plot values, success-rate curves, and operational metrics like the number of false-positive exploration targets generated per square kilometer. Understanding these metrics is essential because a model that looks impressive on paper can still send drilling crews to worthless ground, wasting millions of dollars per campaign. This guide breaks down each metric, explains why it matters, compares competing evaluation approaches, and identifies the practical mistakes that undermine AI-driven exploration programs as of 2026.

## The Direct Answer: Which Metrics Define Rare Earth Targeting Success

**Also worth reading:** [How Do You Evaluate AI Mineral Targeting Systems Before Drilling in 2026?](https://skymineral.com/knowledge/how_do_you_evaluate_ai_mineral_targeting_systems_before_drilling_in_2026.php) · [How Is AI Driven Critical Mineral Targeting Reshaping Global Resource Exploration in 2026?](https://skymineral.com/knowledge/how_is_ai_driven_critical_mineral_targeting_reshaping_global_resource_exploration_in_2026.php) · [How Is Machine Learning Transforming the Discovery of Critical Minerals in 2026?](https://skymineral.com/knowledge/how_is_machine_learning_transforming_the_discovery_of_critical_minerals_in_2026.php)

The single most important metric in machine learning-based rare earth prospectivity mapping is the Area Under the Receiver Operating Characteristic curve, or AUC-ROC. Published ensemble studies on mineral prospectivity mapping under data scarcity — including work appearing in Nature-family journals between 2023 and 2025 — consistently report AUC values between 0.82 and 0.94 for random forest, gradient boosting, and deep learning classifiers applied to REE targets. An AUC of 0.90 means that if you pick one known deposit location and one randomly chosen barren location, the model ranks the deposit higher 90 percent of the time. Anything below roughly 0.75 is generally considered insufficient for committing field budgets in hard-rock REE systems such as carbonatites, peralkaline intrusions, or ion-adsorption clay deposits.

However, AUC alone is misleading for rare earth targeting because deposits are extraordinarily rare events. In a typical 10,000 km² survey area, true REE mineralization might occupy less than 0.1 percent of pixels. A classifier that predicts "no deposit" everywhere achieves 99.9 percent raw accuracy while being completely useless. That is why practitioners rely on precision-recall curves, F1 scores computed at operating thresholds, and the prediction-area (P-A) plot, which plots prediction rate against occupied area. A strong P-A plot shows the model capturing 60–80 percent of known deposits within the top 20–30 percent of ranked prospective area. When evaluating any vendor or internal model, ask specifically for these numbers rather than headline accuracy figures.

## Why Metrics Behave Differently in Rare Earth Exploration Than Other Domains

Rare earth targeting presents a class imbalance problem more extreme than almost any other applied machine learning domain. Antiviral drug discovery campaigns, by comparison, screen libraries where active compounds may represent 0.5–2 percent of candidates; the recent Transformer-based screening effort that identified 23 antiviral candidates against the Bundibugyo Ebola strain worked within a far denser positive class than any REE dataset. In mineral exploration, geologists may have only 15–40 confirmed deposit locations across an entire province, and many of those were found through decades of manual work, introducing severe spatial sampling bias into training labels.

This scarcity forces metric choices that would seem odd elsewhere. Cross-validation must be spatial rather than random — typically block cross-validation with 25–50 km buffer zones between training and test folds — otherwise the model memorizes the spatial footprint of known camps and reports inflated AUC values of 0.95+ that collapse to 0.65–0.75 when tested on truly unseen terrain. Ensemble strategies developed specifically for data-scarce prospectivity mapping address this by combining multiple weak learners and averaging their uncertainty estimates, then reporting not just point predictions but calibrated probability surfaces. A probability of 0.7 from a well-calibrated model should correspond to roughly a 70 percent historical hit rate among areas assigned that score; calibration curves (reliability diagrams) are therefore a metric worth demanding alongside AUC.

## Core Metric Definitions Every Exploration Team Should Know

Beyond AUC, several complementary metrics form the standard evaluation stack. Precision (positive predictive value) tells you what fraction of flagged cells actually contain mineralization — critical because each false positive costs real money in geochemical sampling, geophysics, or drilling. Recall (sensitivity) measures what fraction of true deposits the model catches; missing a world-class deposit is often costlier than chasing ten duds. The F1 score harmonizes both, though it hides threshold sensitivity. The success-rate curve, plotted as cumulative percentage of deposits captured versus percentage of study area explored, remains the industry's favorite visual because executives understand it instantly: a model whose curve rises steeply means you can drill 10 percent of the land and expect to encounter half the deposits.

Distance-based and economic metrics matter too. Mean distance from predicted high-prospectivity zones to nearest known occurrence (often reported in kilometers), target density (high-scoring cells per 100 km²), and expected value calculations that weight each discovery by grade-tonnage assumptions convert statistical performance into boardroom language. Drone-based magnetic and multispectral surveys — such as the published 2023 UAV program over Qullissat on Greenland's Disko Island that built 3D mineral exploration models — feed these pipelines with high-resolution input layers, and the marginal improvement in AUC attributable to adding each data layer (ablation analysis) is itself a metric worth tracking before spending on new acquisition.

## Comparison Table: Classical Statistical Methods Versus Modern ML Approaches

| Feature | Weights-of-Evidence / Logistic Regression | Ensemble Machine Learning (Random Forest, XGBoost, Neural Nets) |
| --- | --- | --- |
| Typical AUC on REE datasets | 0.70–0.82 | 0.82–0.94 |
| Data requirement | Moderate; assumes conditional independence | High; benefits from thousands of labeled cells |
| Handling of non-linear feature interactions | Poor to limited | Strong; captures complex geochemical interactions |
| Interpretability | High; explicit weights per evidence layer | Low to moderate; requires SHAP or permutation analysis |
| Performance under label scarcity (

Canonical: https://skymineral.com/knowledge/what_metrics_matter_most_when_using_machine_learning_for_rare_earth_mineral_targeting.php
Markdown: https://skymineral.com/knowledge/what_metrics_matter_most_when_using_machine_learning_for_rare_earth_mineral_targeting.php/index.md