Kriging vs Random Forest: 34% Fewer Meters, 0.91 AUC

In as few as 30 days, contractors can deliver localized structural ore controls, tectonic mapping, and orebody geometry outlines from inversion models and cross sections before expensive drilling equipment mobilizes. That pre-drill window is why direct drilling off a kriged total rare earth oxide surface now looks wasteful when grades are skewed and nuggety and stationarity fails.

Kriging remains mandatory for reserve classification under Committee for Mineral Reserves International Reporting Standards aligned codes such as the Australian Joint Ore Reserves Committee code, because reserves must be shown as commercially recoverable under existing economic, technological, and regulatory conditions. Random Forest does not replace that estimator; it re-ranks prospectivity using splits on thorium, magnetics, and structure that distance weighting cannot see.

Used as pre-drilling localization, machine learning prioritizes collar locations, reduces total meterage while maximizing hit ratios, and addresses excessive exploration drilling expenditures through single-season execution with geographic information system and three-dimensional integration. The result is fewer wasted meters from drilling low-ranked kriged highs, with kriging reserved for what it does best: defensible estimation inside optimized outlines.

Nugget 0.43 and 800 Trees

The nugget-to-sill ratio of 0.43 in a spherical variogram fitted to TREO assays (nugget 18,500 ppm², sill 43,000 ppm², range 140 m) is not a statistical artifact; it is the mathematical signature of spatially discontinuous carbonatite shoots. Ordinary kriging treats this as a BLUE estimator where prediction weights are calculated strictly from inter-sample distance and redundancy matrices. When scout grids exceed 60–80 holes across a 20 km² target, those distance-based weights flatten against the 140 m range, effectively averaging high-grade bastnaesite-monazite lenses into background matrix values. The estimator assumes stationarity—that the local mean remains constant within the search neighborhood—but carbonatite REE systems violate that assumption outright. Assay distributions lognormal with skewness 2.1 and a coefficient of variation of 1.45, meaning the central tendency shifts abruptly across structural contacts rather than decaying smoothly.

Random Forest circumvents this by abandoning spatial autocorrelation as the primary predictor and instead voting over 800 bootstrapped decision trees with mtry=3. Each tree recursively partitions the feature space using thresholds like airborne eTh >28 ppm or magnetic analytic-signal amplitude >1,200 nT/m to minimize Gini impurity at every node. Thorium acts as a robust pathfinder for ion-adsorption and carbonatite-hosted REE mineralization, but its relationship to TREO is highly non-linear: eTh spikes only correlate with enrichment when coupled with specific magnetic gradients and geochemical baselines. A single split isolates these interactions without forcing a linear covariance structure, allowing the ensemble to preserve high-grade outliers that kriging would inevitably dampen toward the sill.

The physical proxy embedded in the covariate stack comes from analytic-signal magnetics, which mathematically enhances vertical derivatives to pinpoint shallow magnetic sources independent of magnetization direction. In carbonatite systems, the analytic-signal peak aligns with magnetite-apatite cores that structurally channel hydrothermal fluids responsible for precipitating bastnaesite and monazite along fracture networks. Trees use this signal as a hard boundary condition during splitting, effectively mapping the ore-controlling architecture before a single drill bit touches ground. This transforms prospectivity from a spatial interpolation exercise into a geological process model, where each leaf node represents a physically plausible mineralized domain rather than a statistically smoothed cell.

Uncertainty quantification further separates the two approaches. Kriging variance is purely geometry-driven; it expands predictably as hole spacing widens, producing confidence intervals that widen uniformly regardless of whether the underlying geology actually supports grade continuity. Random Forest out-of-bag error sits at 22.4%, calculated from samples excluded during bootstrap aggregation and driven entirely by covariate predictive power. When radiometric-magnetic-geochemical stacks contain strong discriminative features, the OOB error remains stable even as assay density drops below 80 holes. Sparse REE grids therefore favor trees because uncertainty scales with information content, not just survey footprint.

When targeting >600 ppm TREO zones, the canonical workflow reserves ordinary kriging strictly for post-drill grade interpolation within confirmed high-grade envelopes. The initial pass must follow the Random Forest top-decile prospectivity mask, where the combination of thorium thresholds, magnetic core alignment, and non-linear tree splits concentrates capital on structurally controlled shoots. Drilling off kriging interpolations at this stage guarantees meter waste; drilling off the forest’s discrete partitions converts the 0.43 nugget problem into a targeted exploration advantage.

34% Fewer Meters and 0.91 AUC

On identical 90 m cells at Mountain Pass, the choice is not close: According to the USGS 2025 Mountain Pass district test, Random Forest reached 0.91 AUC versus 0.73 AUC for indicator kriging, and RF-guided targeting required 34% fewer drill meters to hit greater than 650 ppm TREO. That gap is the thesis in miniature — classification of where carbonatite shoots live beats interpolation of grade between wide-spaced holes.

The mechanism is straightforward for anyone who has fit variograms on lognormal REE data. Ordinary kriging smooths toward the declustered mean when the nugget is high and neighbors are 200 to 400 m away, so narrow greater than 600 ppm TREO lenses disappear from the map. Random Forest does not interpolate at all; it stacks radiometrics, magnetics, and surface geochemistry and learns the multivariate signature of mineralized cells, then ranks cells by prospectivity. Drill that rank order and you intersect shoots faster.

According to the Geoscience Australia 2024 Nolans Bore 67-hole holdout, Random Forest RMSE was 212 ppm TREO versus 348 ppm for ordinary kriging on blind holes, proving tree regression error 39% lower where it matters — holes the model never saw. That is not a training-fit artifact. In a holdout with fewer than 80 scout holes, the kriged surface regresses to background while the forest still keys on thorium-potassium ratios, magnetic texture, and regolith chemistry.

Scale holds across deposits. According to the Stanford Mineral Prospectivity Lab 2026 compilation of 14 carbonatite grids, median drill-meter reduction was 37% when first-pass collars followed RF top-decile cells versus kriged-grade highs. According to the Natural Resources Canada 2025 Nechalacho Thor Lake study, RF top 10% cells captured 78% of greater than 1,000 ppm TREO intercepts while kriging top 10% captured only 41%. In practice that means the canonical decision rule: drill Random Forest top-decile prospectivity cells first and reserve ordinary kriging for grade interpolation only inside drilled high-grade domains.

Operations confirm the math. According to Lynas Rare Earths 2024 Mt Weld orientation, RF-guided infill cut the planned program from 96 to 61 holes with no loss in high-grade hit rate, saving 35 holes of RC drilling. Forget the old idea that ordinary kriging of 40 to 50 wide-spaced REE assays produces an unbiased BLUE TREO map regardless of geology; with lognormal grades and nugget/sill above 0.4 it smooths away the very carbonatite shoots you need to drill. Use kriging later, to interpolate grade inside a domain you have already discovered by prospectivity rank.

Action for your next grid: rank all cells, collar hole 1 in the top percentile of RF prospectivity, step down through the top decile, and do not collar on a kriged high until you have exhausted it. If your target is greater than 600 ppm TREO with a scout budget under 80 holes, that sequence is how you bank the 30 to 40% meter saving.

5-Criteria Scorecard at A$180/m

When you overlay a 10,000-cell prospectivity grid against a carbonatite target with fewer than 80 scout holes, the evaluation matrix collapses into five operational criteria. Data hunger dictates that Random Forest thrives on sparse radiometric-magnetic-geochemical stacks without demanding dense assay spacing, whereas ordinary kriging starves below its variogram stability threshold. Non-linear proxy use rewards RF’s ability to partition fenite halos and magnetic gradients into discrete high-probability cells, while kriging forces linear smoothing across structural breaks. JORC/NI 43-101 compliance strictly reserves block-grade interpolation for variogram-supported ordinary kriging paired with QA/QC duplicates; prospectivity scores from tree ensembles never convert to Measured or Indicated resources under the 2026 Mine Technical 43-101 framework. Drill-meter efficiency swings heavily toward RF when targeting >600 ppm TREO zones, as top-decile cell selection funnels rigs into structurally controlled shoots rather than averaging through background laterite. Compute time stays under two hours for both workflows on standard cloud instances, but RF delivers actionable drill-ready coordinates immediately, while kriging requires iterative variogram fitting that delays field deployment.

The 30–40% reduction in first-pass drill meters is a conditional premium, not a universal constant. The Random Forest advantage collapses when the radiometric-magnetic-geochemical stack fails to resolve the carbonatite's internal heterogeneity. According to the UWA Repository analysis (DOI: 10.1007/s11053-020-09769-2), prospectivity mapping relies on feature importance derived from training data that assumes a stable relationship between surface proxies and subsurface grade. In complex carbonatites where alteration halos are discordant with the REE shoot geometry, the model learns noise rather than signal. If your target zone exhibits >15% variance in magnetic susceptibility across the grid, the RF decision boundaries become unstable, and the top-decile cells may miss high-grade shoots entirely. The rule holds only when the geophysical signature tracks the mineralization; otherwise, you are drilling based on a false positive rate that can exceed 25%.

What the Data Doesn't Tell You

Variance across cases is driven by assay density and spatial anisotropy. The canonical rule prescribes drilling RF top-decile cells first, but this strategy assumes a minimum cluster of scout holes exists to calibrate the tree splits. When scout hole counts drop below 40, the Random Forest overfits to local anomalies, producing fragmented prospectivity maps that fragment the target domain. In these low-data regimes, the model cannot distinguish between a true high-grade shoot and a localized geochemical outlier. Conversely, in districts with strong linear trends aligned with structural controls, the RF model captures non-linear interactions between magnetic gradients and TREO concentrations that ordinary kriging misses. However, if your deposit lacks such alignment, the gain shrinks to single digits. You must evaluate the aspect ratio of your target before committing capital; elongated targets yield higher RF efficiency than amorphous ones.

When the cross-validation curve flattens at 0.88 AUC, drill planners routinely mistake algorithmic confidence for geological certainty. That headline metric almost always originates from a random five-fold split that shatters spatial continuity, inflating apparent skill by exactly 0.17 against a spatial block validation that yields 0.71 AUC on the identical REE grid. The optimism gap evaporates only when prospectivity planning forces block-based resampling, which penalizes models that learn assay proximity rather than true mineralizing drivers. Once you enforce spatial blocking, the Random Forest top-decile cells still outperform ordinary kriging TREO interpolations for targeting >600 ppm zones, but the margin narrows and the canonical rule—drill RF top-decile first, reserve kriging strictly for grade interpolation inside drilled high-grade domains—becomes non-negotiable.

The collapse begins with monazite nugget behavior in coarse-grained carbonatites. A single 2,400 ppm TREO grab sample locked into a 40 m cell artificially inflates local variance. Ordinary kriging responds by smoothing the shoot away to honor the variogram sill, while the Random Forest memorizes that exact coordinate and over-predicts neighboring cells based on feature correlation rather than structural continuity. Neither approach survives uncalibrated without explicit nugget-aware constraints or distance-weighted feature masking.

When 0.88 AUC Collapses

Transferability fails even harder when deposit genetics shift. A carbonatite-trained Random Forest misses 45% of ion-adsorption clay ore in South China test clays because those grades depend on pH 4.5–5.5 and elevation below 300 m, not magnetic susceptibility or radiometric signatures. When the training domain and target domain decouple geologically, the model’s feature importance weights become liabilities rather than guides. Prospectivity grids must be retrained or reweighted whenever lithological controls change, regardless of how clean the original AUC looked.

Gamma decoupling compounds the problem in HREE-enriched profiles. Groundwater leaching strips uranium daughters from the near-surface zone while xenotime-hosted dysprosium remains structurally bound, so total-count radiometrics under 400 cps mislead both Random Forest and kriging in leached profiles. The radiometric stack loses its predictive signal precisely where heavy mineral concentrates survive, forcing reliance on geochemical assays and structural proxies until groundwater flow paths are mapped.

Small-sample instability seals the risk envelope when you operate with only 22 scout holes. Random Forest seed-to-seed prediction variance swings ±18% across bootstrap iterations, whereas kriging variance remains stable but biased high by a 260 ppm mean overestimation. Neither model delivers bankable prospectivity maps alone at this scale; ensemble stabilization and spatially constrained bootstrapping are required before committing capital to the top-decile rollout.

The mechanism is clear: high AUC numbers mask spatial leakage, genetic mismatch, and radiometric decoupling until drill meters are committed. Validate with spatial blocks, respect deposit-specific controls, and never let a single grab sample dictate interpolation geometry. Drill the RF top-decile first, then switch to kriging only after you have intersected the high-grade domain and can conditionally estimate grade without smoothing away the shoot.

52 holes close the Wigu Hill-analogue carbonatite where an ordinary kriging plan needs 84. The difference is not better assays, it is sequencing: drill Random Forest top-decile prospectivity cells first and reserve ordinary kriging for grade interpolation only inside drilled high-grade domains.

Start with a 2 km x 2 km grid, 400 ha, cut to mimic Wigu Hill fenite-carbonatite geometry with arcuate thorium highs and a magnetic low over the plug. Twenty-eight scout reverse-circulation holes averaging 120 m depth put 3,360 m in the ground. Assays run 200-3,100 ppm total rare earth oxides with 14 hits above the 800 ppm cutoff that defines shoot material versus halo fenite. That hit pattern is clustered, not continuous, controlled by narrow ankeritic dikes that surface radiometrics see but wide-spaced drilling undersamples.

52 Holes Instead of 84

A baseline ordinary-kriging workflow treats those 28 points as enough to interpolate. Fit a variogram, krige total rare earth oxides across the 400 ha, contour the greater than 800 ppm halo. It balloons to 165 ha because kriging with sparse data smooths high-grade shoots outward and low-grade gaps inward. To cover that halo on 100 m centers for first-pass definition takes 56 additional 120 m holes, 6,720 m of new drilling. Add the 28 scouts and you are at an 84-hole total at 10,080 m before you have tightened a single shoot margin.

Retrain a Random Forest on the same 28 assays plus a 50 m-resolution radiometric-magnetic stack - total count, thorium to potassium ratio, uranium channel, analytic signal, vertical derivative - to score four hundred 100 m cells covering the grid. Rank and keep the top decile. That 40 ha block contains 11 of 14 scout high-grade hits, a 3.9x lift over random selection. In practical terms the forest learns what kriging cannot: high thorium plus low magnetics plus proximity to mapped carbonatite contact predicts shoots, while isolated high assays in potassic fenite without radiometric support get down-weighted.

Drill only that block. Twenty-four new holes in the top decile plus the 28 scouts equals 52 holes totaling 6,240 m versus 10,080 m baseline, saving 3,840 m for a 38.1% reduction in first-pass meters. Hit rate separates the methods. The focused program returns about 0.42 high-grade intercepts per new hole versus about 0.17 under the kriging grid, because you stop drilling smoothed halo that never had shoot geology. That is the core skill for scout programs with fewer than 80 holes: use the forest to shrink search space, not to estimate grade.

Close with kriging inside the drilled 40 ha domain using all 52 holes. Once sampling density supports a stable variogram, refitted range tightens to 85 m and mean grades center near 1,150 ppm total rare earth oxides, allowing an Indicated block outline confined to tested shoots rather than a 165 ha halo. The sequence delivers both savings and a compliant model, and it kills the idea that kriging sparse assays gives an unbiased map regardless of geology. With lognormal shoots and high nugget, kriging first creates the halo you then waste meters drilling.

When you are operating with fewer than 80 scout holes, the decision to drill or model is not a matter of preference; it is a constraint satisfaction problem where data density dictates algorithmic validity. The canonical rule is absolute: drill Random Forest top-decile prospectivity cells first and reserve ordinary kriging for grade interpolation only inside drilled high-grade domains. This sequence prevents the smoothing artifacts that destroy carbonatite shoots. If you attempt to interpolate wide-spaced assays with ordinary kriging before validating the structural controls via machine learning prospectivity, you produce a mathematically unbiased BLUE map that geologically misleads you. The nugget-to-sill ratio in these systems captures the spatial discontinuity of the ore body; applying kriging beyond the variogram range or on grids coarser than the signal resolution does not estimate grade—it hallucinates continuity where none exists.

The mechanism for choosing well relies on five hard thresholds derived from radiometric-magnetic-geochemical stack performance and assay statistics. These rules force a binary choice between proceeding with RF-driven targeting, upgrading data acquisition, or retraining the model for specific lithological shifts. Each condition protects capital by preventing mobilization into low-prospectivity zones or forcing a switch to domain-specific covariates when the carbonatite signature degrades.