REE Intelligence Platform

Comprehensive AI-powered rare earth element analysis — from data exploration to drill target recommendations
Dataset: 5,763 samples across 1,254 drill holes · Coverage: 13.9 km (N–S) × 14.4 km (E–W)
1 Analysis Overview
What AI/ML techniques were applied, what data was used, what was predicted, and how it helps decision-making.

This dashboard presents the results of a multi-stage AI/ML analysis of 1,254 drill holes (5,763 geochemical samples) across the project area. The system applies four complementary machine learning models to transform raw drilling data into actionable exploration intelligence — predicting grades at unsampled locations and depths, identifying geochemical zones, and ranking drill targets by potential value.

AI/ML Pipeline
1
Feature Engineering
Raw assay data (16 REE elements) transformed into spatial and geochemical features: compositional ratios, anomaly indices, neighbourhood statistics, and distance metrics.
39 features
2
Surface Grade Model
XGBoost + Random Forest ensemble predicts surface TREE grade at every hole and across a spatial heatmap grid. Validated by spatially separated cross-validation.
R² = 0.58
3
Depth Extrapolation
Variance-scaled gradient boosting trained on 215 multi-depth holes with Gaussian-kernel surface anchoring. Predicts grade at 14 depths (5–500 m) for 1,039 surface-only holes.
±47 ppm
4
Zone & Target ID
Gaussian Mixture clustering identifies geochemical zones. Composite scoring ranks all holes by predicted grade, confidence, proximity, and surface grade for drill priority.
15 top targets
Inputs
Drill Hole Assays
5,763 geochemical samples from 1,254 holes — 16 rare earth elements (La through Lu plus Y), with location coordinates (Easting, Northing) and depth intervals.
Multi-Depth Profiles
215 holes drilled at multiple depths provide the training data for depth extrapolation — the system learns how grade changes underground from these measured profiles.
Spatial Context
Hole coordinates enable neighbourhood analysis — the model computes what nearby holes show, how far each hole is from known high-grade areas, and local grade trends.
Outputs
Predicted Surface Grade
AI-estimated TREE concentration at every hole and across a spatial heatmap grid — accuracy of ±78 ppm (R²=0.58).
Predicted Depth Profiles
Grade forecasts at 14 depths (5–500 m) for 1,039 single-interval holes, with quantified uncertainty envelopes — accuracy of ±47 ppm.
Drill Priority Ranking
Every scored hole ranked 0–100% by composite worth (predicted grade, confidence, spatial coverage, surface grade) — 15 holes score above 80%.
What's in This Dashboard
2
Executive Summary
Key metrics at a glance — grade, prediction accuracy, targets, and coverage.
3
Spatial Map
Interactive map with 6 modes: measured grade, predicted grade + heatmap, HREE%, Dy, zones, and drill priority.
4
Geochemical Zones
AI-identified zones with boundary confidence ratings for resource planning.
5
Grade Analysis
Element enrichment, HREE distributions, and economic indicators.
6
Prediction Drivers
Which geological and spatial factors most influence the AI predictions.
7
Depth Intelligence
Predicted grade at depth, project-wide trends, and multi-hole comparison.
8
Drill Recommendations
Ranked drill targets with expected grade, recommended depth, and confidence.
9
Validation
How the models were tested — spatially separated cross-validation ensures trustworthy predictions.
10
3D Digital Twin
Interactive 3D prototype — Leapfrog-style grade shells, drill holes, and priority targets in three dimensions.
Practical Value for Decision-Makers
🎯 Targeted Drilling — Reduce Exploration Cost
Instead of drilling blind, the AI identifies 15 high-priority targets with predicted grade, recommended depth, and confidence level. This focuses capital on the locations most likely to yield economic mineralisation, potentially reducing the number of speculative holes needed.
📈 Grade Prediction — Informed Resource Estimation
Surface grade predictions (±78 ppm accuracy) and spatial heatmaps allow resource planners to estimate TREE distribution across the project area before additional drilling — bridging gaps between sparsely drilled locations.
⚒ Depth Forecasting — Optimise Drill Depth
For 1,039 holes with only surface data, the system predicts what grades to expect at 5–500 m depth with quantified uncertainty. The recommended drill depth identifies the shallowest depth where grade peaks — avoiding unnecessarily deep (costly) drilling.
🚀 Scalable Foundation — From Prototype to Platform
This analysis demonstrates proven AI/ML value on real data. The same models can be retrained automatically as new assays arrive, forming the foundation for a real-time exploration intelligence platform with 3D visualisation and continuous learning.
Data Context
Data coverage:
82.9% of holes (1,039) are single-interval with only surface data (0–5 m). The 215 multi-interval holes (drilled to median 100 m) provide the training data for depth predictions. The AI extends these measured depth patterns to all 1,039 untested locations using spatial proximity and surface geochemistry.
HREE value opportunity:
346 samples exceed 25% HREE/TREE — Dy and Tb are 50–200× more valuable per kg than light REEs (La/Ce). The Zones view on the spatial map highlights where these high-value geochemical fingerprints cluster, helping prioritise targets with both depth potential and HREE enrichment.
2 Executive Summary
Key numbers at a glance — combining spatial analysis, geochemical characterisation, and predictive modelling across the entire project area.
Avg TREE Grade
248
ppm across all holes
Peak Grade
782
ppm — D48
Dy Enrichment
5.5×
high-grade vs low-grade ratio
HREE Premium Samples
346
samples ≥ 25% HREE/TREE
Grade Prediction
Moderate
within ±78 ppm on average
Depth Prediction
High
within ±47 ppm on average
Top Drill Targets
15
scoring >80% drill priority
Data Coverage
83%
holes with only surface data (1039 of 1254)
3 Project Overview — Spatial Map
Interactive map of all 1,254 drill holes. Switch modes to view average TREE grade, AI-predicted surface grade, heavy-REE content, dysprosium, refined geochemical zones with boundary confidence, or drill priority scoring. Hover any hole for detailed information including surface measured vs. predicted comparison.
INSIGHT: 468 of 1254 holes (37%) exceed 300 ppm TREE. Median grade: 200 ppm; peak: 782 ppm (D48). High-grade holes concentrate in the central-east area.
4 Geochemical Zones
AI-identified zones based on REE composition and spatial location. Boundary confidence ratings indicate how stable each zone definition is — higher confidence means the boundary is well-defined and reliable for planning.
Zone A
Holes591
Avg TREE259.6 ppm
Avg HREE%20.0%
Area19.5 km²
Boundary confidence✓ High (99.2%)
Compact spatial zone
High confidence — zone boundary is well-defined and reliable for resource planning
Zone B
Holes214
Avg TREE221.3 ppm
Avg HREE%18.5%
Area40.4 km²
Boundary confidence✓ High (88.9%)
Compact spatial zone
High confidence — zone boundary is well-defined and reliable for resource planning
Zone C
Holes30
Avg TREE190.8 ppm
Avg HREE%17.1%
Area2.3 km²
Boundary confidence● Moderate (83.9%)
Compact spatial zone
Moderate confidence — zone boundary is generally stable; consider infill drilling at margins
Zone D
Holes419
Avg TREE249.2 ppm
Avg HREE%18.7%
Area20.4 km²
Boundary confidence✓ High (97.3%)
Compact spatial zone
High confidence — zone boundary is well-defined and reliable for resource planning
Action guidance: Zones rated High confidence have stable boundaries suitable for resource estimation. Moderate confidence zones could benefit from infill drilling at margins. Low confidence zones need additional delineation drilling.
5 Grade Analysis
Element enrichment patterns, grade distributions, and economic indicators across the project.
Element enrichment factors
High-grade vs low-grade sample contrast for each rare earth element.
Tb
14.50×
Gd
5.60×
Tm
5.50×
Dy
5.50×
Lu
5.50×
Er
5.30×
Y
5.00×
Ho
5.00×
Sm
5.00×
Nd
4.90×
Yb
4.80×
Pr
4.60×
Sc
4.30×
Ce
4.30×
Eu
4.30×
La
3.90×
Key finding: Tb shows 14.5× enrichment — the strongest zonation signal. Tb and Dy spatial patterns should guide resource boundary delineation.
HREE ratio distribution
HREE/TREE% histogram. Bar colours: Low (<18%) · Moderate (18–22%) · Elevated (22–25%) · Premium (≥25%).
Key finding: 346 premium samples (≥25% HREE/TREE) carry disproportionate economic value — Dy and Tb are 50–200× more valuable per kg than La/Ce. Mean HREE/TREE across the project: 19.3%. The right tail of this distribution identifies the highest-value targets.
TREE by drilling method
Mean TREE by sample type. Grade differences between methods should be considered in resource models.
Key finding: Reverse circulation (3,592 samples) yields the highest mean grade at 313 ppm, while Trench averages 188 ppm — a 125 ppm spread. This systematic difference must be accounted for in resource models to avoid bias.
Dysprosium vs TREE
Each point coloured by HREE%. R²=0.91 — Dy tracks TREE almost perfectly, simplifying exploration targeting.
Key finding: Dy and TREE are highly correlated (R²=0.91), meaning Dy can serve as a reliable proxy for total REE grade. This simplifies field-level targeting — measuring Dy alone gives a strong indication of overall TREE content. Samples with elevated HREE% (orange/red points) cluster at higher grades.
6 What Drives the Predictions
Understanding which geological and spatial factors most influence the grade estimates — for both surface and depth predictions.
Surface grade — key factors
Which characteristics most influence predicted surface TREE grade.
REE element composition
Grade of 3 nearest sampled holes
Europium anomaly (Eu/Eu*)
Grade of 5 nearest sampled holes
Light-to-heavy REE enrichment ratio
Distance to known high-grade holes
Key finding: Strongest predictors are REE composition, grades of nearby holes, and geochemical signatures (Europium anomaly, light-to-heavy REE ratio).
Depth grade — key factors
Which factors most influence predicted grade at depth (different from surface).
Europium anomaly at depth (Eu/Eu*)
Depth behaviour of nearest deep holes
Heavy REE proportion
Measured surface TREE grade
REE element composition
Heavy-to-middle REE ratio at depth
Light-to-heavy REE ratio at depth
Cerium anomaly at depth (Ce/Ce*)
Key finding: Depth predictions are driven by Europium anomaly at depth, depth behaviour of nearest deep holes, and measured surface grade.
What is the Europium anomaly (Eu/Eu*) and why does it matter?

The concept

Europium (Eu) behaves differently from its periodic-table neighbours Samarium (Sm) and Gadolinium (Gd). The Europium anomaly (Eu/Eu*) measures how much measured Eu deviates from what the neighbours predict.
Eu/Eu* = Eu ÷ √(Sm × Gd)
Reading the number: 1.0 = as expected. Below 1.0 = Eu depleted (negative anomaly). Above 1.0 = Eu enriched.

Why it matters for this project

Eu/Eu* correlates negatively with TREE grade (r = -0.31). Lower Eu/Eu* → higher REE grades. Median across all samples: 0.258.

Calculation example

1
Look up Samarium (Sm) and Gadolinium (Gd). For hole D48: Sm = 21.5 ppm, Gd = 17.55 ppm
2
Calculate expected Europium (Eu*) = √(Sm × Gd) = √(21.5 × 17.55) = 19.42 ppm
3
Divide measured Eu by expected: Eu/Eu* = 4.94 ÷ 19.42 = 0.254
4
A value of 0.254 (well below 1.0) indicates a strong negative Europium anomaly — consistent with high REE grades (781 ppm TREE).

High-grade vs moderate-grade comparison

HoleTREESmEuGdEu*Eu/Eu*Grade
D4878121.54.9417.5519.420.254high-grade
01-06-2012-02B1454.091.243.463.760.33moderate-grade
7 Depth Intelligence
Predicted grade profiles at depth, project-wide depth trends, and measured multi-hole depth comparison.
Predicted grade at depth — What to expect if you drill deeper
For each target hole, the system predicts TREE grade at 14 depths (5–500 m). The solid line is the best estimate; the shaded area is the 80% confidence range. Select a hole to view its profile.
How to read: Solid teal = predicted grade (P50). Shaded band = uncertainty. Diamond = measured surface grade. Coloured lines = real measured profiles from nearest deep holes for comparison. Red dashed = 300 ppm high-grade reference line.
How this was calculated: The AI depth model learns depth-grade patterns from 215 multi-interval holes. For each target, the model uses surface grade, geochemistry, spatial proximity to deep holes, and depth features to predict grade at 14 depths (5–500 m). Accuracy by depth: 0–50 m: ±40 ppm, 50–150 m: ±47 ppm, 150–300 m: ±59 ppm, 300–500 m: ±66 ppm.
Grade outlook by depth
Project-wide average grade and probability of exceeding economic thresholds at each depth level.
5m
Avg 235 ppm
32% chance of >300 ppm
21% chance of >400 ppm
10m
Avg 236 ppm
32% chance of >300 ppm
22% chance of >400 ppm
20m
Avg 238 ppm
33% chance of >300 ppm
22% chance of >400 ppm
30m
Avg 241 ppm
33% chance of >300 ppm
21% chance of >400 ppm
50m
Avg 248 ppm
35% chance of >300 ppm
22% chance of >400 ppm
75m
Avg 246 ppm
37% chance of >300 ppm
20% chance of >400 ppm
100m
Avg 257 ppm
39% chance of >300 ppm
20% chance of >400 ppm
150m
Avg 292 ppm
41% chance of >300 ppm
25% chance of >400 ppm
200m
Avg 288 ppm
39% chance of >300 ppm
25% chance of >400 ppm
250m
Avg 310 ppm
46% chance of >300 ppm
26% chance of >400 ppm
300m
Avg 283 ppm
39% chance of >300 ppm
21% chance of >400 ppm
350m
Avg 264 ppm
36% chance of >300 ppm
16% chance of >400 ppm
400m
Avg 247 ppm
35% chance of >300 ppm
14% chance of >400 ppm
500m
Avg 220 ppm
28% chance of >300 ppm
7% chance of >400 ppm
Key finding: Best chance of exceeding 300 ppm is at 250m (46% probability). Important context: the predicted averages here are for the 1,039 single-interval holes (surface mean 235 ppm), while measured depth data comes from 215 multi-interval holes (surface mean 318 ppm) — an 83 ppm gap because historically, higher-grade locations were selected for deeper drilling. This population difference explains why predicted shallow grades appear lower than measured shallow grades: the model is predicting for a lower-grade population, not underestimating. Use the individual hole depth profiles above for site-specific drilling decisions.
How this was calculated: The AI depth model (gradient boosting trained on 215 multi-interval holes) predicted TREE grade at each depth level for all 1,039 single-interval holes. The average ppm at each depth is the mean of predicted P50 values across all 1,039 holes. The chance of >300 ppm is the percentage whose predicted P50 exceeds 300 ppm; likewise for >400 ppm. Note on populations: the averages above reflect the 1,039 single-interval holes (avg surface grade 235 ppm). The measured depth statistics in the training data panel below come from 215 multi-interval holes (avg surface grade 318 ppm). The ~83 ppm gap exists because higher-grade locations were historically selected for deeper drilling — this is a sampling bias in the data, not a model error.
Training data — measured multi-interval holes
Summary of the 215 multi-interval drill holes (4,698 samples) that the AI depth model learns from. These measured depth profiles are the ground truth for all predictions.
215
multi-interval holes
4,698
total samples
100m
median max depth
641m
deepest hole
Depth profiles — top 6 deep holes (10 m bins)
Training data depth reach
≥50m
99%
213 holes
≥100m
73%
156 holes
≥150m
9%
20 holes
≥200m
8%
18 holes
≥300m
6%
12 holes
≥500m
1%
2 holes
Measured TREE grade by depth band (mean ppm)
0–25m
322
ppm (n=759)
25–50m
322
ppm (n=1,028)
50–100m
316
ppm (n=1,872)
100–150m
286
ppm (n=233)
150–200m
233
ppm (n=217)
200–300m
248
ppm (n=327)
300–500m
259
ppm (n=206)
Key finding: Grade is stable through 100 m (~319 ppm) then drops ~22% by 150 m. Below 150 m, only 9% of training holes have data — predictions at greater depths rely on fewer measured profiles and carry wider uncertainty. Among 6 deep holes, LP2011-02 maintains elevated grade (>200 ppm) down to 630.0m. Across all training holes, 47% show increasing grade with depth and 53% decreasing. Sampling bias note: these 215 multi-interval holes have a higher average surface grade (318 ppm) than the 1,039 single-interval holes (235 ppm), because historically higher-grade locations were selected for deeper drilling. The depth model accounts for this when making predictions.
How this was calculated: All statistics above are computed directly from the 4,698 measured assay samples across 215 multi-interval drill holes — no model predictions are involved. The depth reach shows what percentage of training holes were drilled to at least each depth threshold. The TREE by depth band shows the mean measured TREE concentration within each depth interval. These measured patterns are the ground truth that the AI depth model learns from to predict grade at 1,039 untested single-interval locations.
8 Recommended Actions — Where to Drill Next
Drill targets ranked by composite depth potential score — combining predicted grade, depth value, data confidence, spatial coverage, and surface grade into a single actionable metric.
Recommended drill targets
The Recommended Drill Depth is the shallowest depth where predicted grade reaches within 2% of the peak — avoiding unnecessarily deep drilling when grade is already near maximum at shallower depths.
#Hole IDSurface Grade
(ppm)		Depth Potential
Score		Recommended
Drill Depth (m)		Expected Grade
at Depth		Uncertainty
Range (ppm)		Dist. to Nearest
Deep Hole
1SS0420-15503★ 85%300m560110 (tight)128m
2RE0420-175508★ 84%150m656114 (tight)307m
3LP00163481★ 84%300m522116 (tight)21m
4LP00153483★ 83%250m584128 (moderate)67m
5SS0420-37448★ 82%250m54191 (tight)279m
612-7-10-14390★ 82%350m48596 (tight)29m
7SS0420-39460★ 82%150m561118 (tight)188m
812-8-10-13443★ 82%250m494104 (tight)106m
9SS0420-34409★ 81%200m48989 (tight)132m
1011-3-10-17421★ 81%150m491106 (tight)64m
1110-6-10-20374★ 81%350m486104 (tight)36m
1212-8-10-12465★ 81%250m531120 (tight)129m
1311-3-10-14446★ 80%250m612142 (moderate)63m
14LP00165476★ 80%300m540120 (tight)200m
15RE0420-129467★ 80%150m561119 (tight)270m
16LP00149426● 80%250m493113 (tight)108m
172-13-11-02571● 80%150m627150 (moderate)137m
1811-3-10-24350● 79%250m45487 (tight)53m
19SS0420-38449● 79%350m498108 (tight)250m
2012-8-10-08454● 79%250m499112 (tight)222m
How to use this table: Focus on ★ scores above 80% — these combine high predicted grades, good confidence, and proximity to proven deep holes. Tight uncertainty (bottom third of all holes) = most confident predictions. Wide (top quarter) = more risk. Closer distance to deep holes = more reliable predictions.
Key finding: The top 20 targets have an average recommended drill depth of 245m with an average expected grade of 534 ppm. Of the 1,039 scored holes, 15 (1%) score above 80% — these should be the first priority for drilling.
How this was calculated
Each of the 1,039 single-interval holes receives a Depth Potential Score (0–100%) computed from five weighted factors:
  • Predicted grade at depth — the AI-estimated TREE concentration at the optimal depth, based on patterns learned from 215 measured multi-depth holes (heaviest weight)
  • Prediction confidence — how tight the uncertainty envelope is around the prediction; tighter = more trustworthy
  • Depth value — rewards targets where predicted grade peaks at meaningful depth below the existing surface sample; shallow-peaking holes are penalised since drilling would not provide significant new information
  • Spatial coverage — proximity to proven deep holes; closer = more reliable depth predictions
  • Surface grade — measured surface TREE grade as a baseline indicator
The Recommended Drill Depth is the shallowest depth where predicted grade reaches within 2% of the peak — avoiding unnecessarily deep (and costly) drilling when grade is already near maximum at shallower depths. The Uncertainty Range is the average width of the 80% confidence band across all predicted depths.
9 How Confident Are We?
Validation methodology and accuracy metrics — how the system was tested to ensure predictions are trustworthy.
Surface Grade Predictions
Accuracy levelModerate (58% of variation explained)
Average error±78 ppm TREE
Accuracy by regionR² ranges 0.25–0.58 across 3 spatial folds — accuracy varies by area
Test methodSpatially separated (3 rounds) — each test area held completely apart from training data
Training data1,254 holes (5,763 samples) across 13.9 × 14.4 km
What this means: A predicted grade of 300 ppm is most likely between 222 and 378 ppm actual. Accuracy varies by location — some areas of the project are better constrained by nearby data than others, reflected in the R² range of 0.25–0.58 across spatial folds.
Depth Grade Predictions
Accuracy levelHigh (74% of variation explained)
Average error±47 ppm TREE
Uncertainty reliability81% — predicted range captures true grade 81% of the time
Accuracy by depth0–50 m: ±40 ppm, 50–150 m: ±47 ppm, 150–300 m: ±59 ppm, 300–500 m: ±66 ppm
Test method10-fold grouped-hole CV — holes grouped so no test hole's data leaks into training
Training data215 multi-interval holes (4,698 samples), median depth 100 m, deepest 641 m
What this means: A predicted depth grade of 300 ppm is most likely between 253 and 347 ppm actual. Predictions are more accurate at shallow depths (±40 ppm above 50 m) and less certain deeper (±66 ppm beyond 300 m) where only 6% of training holes have data. The 215 training holes have a higher average surface grade (318 ppm) than the 1,039 predicted holes (235 ppm) due to historical drilling bias — the model accounts for this difference.
How was this validated?
The system was tested by hiding data from known holes and checking whether it could predict their grades correctly — the gold standard for validating prediction models. Surface predictions used spatially separated test areas (3 independent rounds). Depth predictions used 10-fold grouped-hole CV across 215 multi-interval holes — holes are grouped so that no test hole's data leaks into training, ensuring honest accuracy estimates.
10 3D Digital Twin — Deposit Visualisation
Interactive 3D prototype of the deposit based on predicted depth profiles, drill priority targets, and geochemical domains. Rotate, zoom, and pan to explore the subsurface. This is a preview of a future real-time 3D digital twin platform.
Interactive 3D deposit model
Three views of the project area in 3D. Click the tabs to switch between views. Use mouse to orbit (drag), zoom (scroll), and pan (right-drag). Click legend items to toggle layers on/off.
1,254
drill holes in 3D model
4,698
measured samples (ground truth)
15
high-priority targets
500m
max prediction depth
How to use: Grade Shells shows Leapfrog-style nested isosurface shells at 100, 200, and 300 ppm TREE for both measured (saturated colours) and predicted (lighter, pastel tones) data, with domain boundaries and all drill holes. Where measured and predicted shells overlap, the model is reliable; where they diverge, additional sampling may be warranted. Toggle any layer on/off via the legend. Drill Planning displays top priority targets with predicted grade-depth curves in 3D.
How this was calculated: Grade shells are extracted using the marching cubes algorithm on a 200m × 200m × 25m interpolated grid. Interpolation uses linear interpolation within the data convex hull and Gaussian smoothing (σ=2.0). Three nested isosurfaces are extracted at 100, 200, and 300 ppm, rendered as smooth Phong-lit triangle meshes with per-vertex normals. Measured shells use 4,698 depth assays plus surface samples from 215 multi-interval holes (Zones A–D). Predicted shells use AI depth profiles from 1,039 single-interval holes — note that all single-interval holes fall within Zone B in the clustering, so predicted shells only cover the Zone B spatial extent and should not be directly compared with Zone A/C/D measured shells. Domain boundaries are shown as dashed lines at the surface elevation.
11 Future Platform Capabilities
The analysis presented in this dashboard demonstrates the practical value of AI/ML for REE exploration. Building on these proven foundations, a comprehensive platform could deliver the following capabilities.

⚙ Real-Time AI Training Pipeline

Automatically retrain models as new drill results arrive. The system learns from each new hole, continuously improving predictions.

  • Auto-ingest new assay data
  • Incremental model updates
  • Live accuracy monitoring
  • Automated model versioning

📄 AI-Powered Reporting

Generate professional geological reports using AI that understands your data. Ask questions in plain English and receive data-backed recommendations.

  • Natural language queries
  • Auto-generated technical reports
  • Executive summary generation
  • Anomaly alerts & explanations

📦 3D Digital Twin

A living 3D model of the deposit that updates in real-time. Visualize grade distribution, drill targets, and resource boundaries interactively.

  • Real-time 3D grade visualization
  • Interactive drill planning
  • Block model integration
  • What-if scenario modelling

🌐 Integrated Web Platform

Secure, cloud-hosted platform accessible from any browser. Role-based access for geologists, mine planners, and executives.

  • Cloud-hosted, always available
  • Role-based dashboards
  • API integration with mine systems
  • Mobile-responsive design