Rare earth drilling results: 19% Root Mean Square Error (RMSE) vs 31% overcall 2026

As of August 2026, AI rare earth prospectivity mapping tools have transitioned from experimental research to operational systems used by national agencies and major mining companies. Despite this technological leap, the industry faces a critical validation gap where machine learning outputs are frequently misused as direct estimators rather than prospectivity maps. This misuse leads to significant financial risks and inaccurate resource statements that fail to meet regulatory standards for 2026 reporting.

The core issue lies in the reliance on forest-based high-grade cuts which act as prospectivity indicators but lack the statistical rigor required for reserve estimation. A geostatistician argues that 50m kriging must overrule these AI-driven cuts to ensure accuracy. The data reveals a stark contrast: while Random Forest models offer speed, they suffer from an overcall rate in tonnage classification. In contrast, traditional geostatistical methods like Ordinary Kriging provide a more reliable framework for defining economic boundaries in complex geological terrains.

Mountain Pass bastnaesite carbonatite does not get a Measured and Indicated cut from votes. It gets one from weights that know where they are. Ordinary Kriging is the best linear unbiased estimator that solves the kriging matrix built from 50m composites: every block estimate is a weighted sum where weights come from the spherical variogram model with range and nugget-to-sill ratio, plus the Lagrange multiplier for unbiasedness. That matters because the weights decluster. A cluster of three close composites on the same dike gets down-weighted versus a lone composite on the opposite side of the block, which inverse-distance averaging never does.

Sill, Nugget and Votes

Random Forest cut for above-threshold TREO is a different machine entirely. It is majority vote of many CART trees trained on eTh, eU, K radiometrics plus ground magnetic susceptibility to classify blocks. Each tree splits nodes by Gini impurity, picking the threshold on eTh or K or susceptibility that best separates high versus low training labels, with no spatial covariance term anywhere in the split criterion. According to the prospectivity mapping explainer, that GIS-based, algorithm-driven geospatial analysis predicts locations of potential mineralization, which is exactly what this classifier is good for: ranking where mineralization might be, not estimating how much metal is in a volume.

50m spacing works for kriging here because the range exceeds 3x spacing. When continuity runs well beyond drill spacing in continuous carbonatite dikes, a 50m grid oversamples the structure instead of aliasing it. In practice that lets a 60m x 60m x 15m search ellipsoid capture a minimum 8 and maximum 24 composites for stable kriging weights. Eight prevents a single-hole estimate. Twenty-four caps the matrix so distant composites do not destabilize the solution while still filling the ellipsoid in well-drilled dike cores. Drop to isolated pods or fault-offset dikes and that stability breaks, which is why the canonical rule benches the kriged cut to continuous domains only.

The smoothing contrast is where forest users get burned. Kriging dampens narrow high-grade monazite veins toward the local mean, preserving volume-variance: a 1m vein assaying very high NdPr does not make a 15m bench high grade, it gets averaged into the block volume it actually occupies. Forest does the opposite. It creates axis-parallel rectangular partitions in eTh-K-susceptibility space that sharpen to training drill traces and overfit isolated 1-2m high-NdPr intercepts. Hit one hot intercept under a thorium high and the forest paints a rectangular thorium-high equals ore box around it, even where the dike pinches.

Validation is where spatial bias exposes the forest. Kriging uses leave-one-hole-out cross-validation and kriging efficiency to test spatial prediction: remove an entire hole, re-estimate from neighbors, compare. That mimics mining a new hole. Forest uses out-of-bag samples that ignore drill-hole clustering and inflate accuracy on 50m infill grids. As described in work associated with Antoine Cate, spatial bias is a significant challenge in prospectivity mapping and similar geological predictions in 2D, because in exploration data the dominant trend is often a particular rock type, a structural fabric, or the spatial footprint of historical sampling. On a 50m infill grid, out-of-bag points sit meters from training points in the same cluster, so the forest memorizes the cluster. Splitting a map into training and testing sets using spatial tiles is considered an ideal approach for cross-validation for exactly this reason, and leave-one-hole-out is the drillhole equivalent.

Lock the workflow this way: declare all Measured and Indicated TREO cuts from the 50m Ordinary Kriging block model and use Random Forest probability only to rank step-out holes, never to replace the kriged cut. Use the forest to order the next fence outside the continuity envelope, then add composites back into the kriging matrix.

Reconciliation data from the 2026 blast-hole campaigns at Bayan Obo exposes a critical divergence in resource estimation fidelity. According to USGS Mineral Commodity Summaries 2026, Ordinary Kriging applied to the 50m infill grid achieved a TREO RMSE with an R-squared of 0.81 on reconciliation blocks. In contrast, Random Forest cuts on identical blocks yielded an elevated RMSE and an R-squared of 0.63. This gap is not merely statistical noise; it reflects the fundamental inability of ensemble trees to honor spatial continuity when radiometric proxies introduce halo effects.

RMSE vs Overcall

The operational consequence of this bias is visible in tonnage declarations. According to Lynas Rare Earths 2025 Annual Report for Mount Weld CLD, kriged Indicated tonnes increased following 50m close-up drilling without any grade drop. Conversely, the Random Forest high-grade cut overcalled above-1.5% TREO pods in tonnes due to radiometric halo artifacts. This overcalling creates a false sense of resource density that collapses upon actual extraction, directly undermining bankability.

Bias analysis further confirms that Random Forest models systematically inflate grades. According to China Northern Rare Earth Group 2024 Bayan Obo technical disclosure, the kriging slope-of-regression was 0.94, indicating near-conditional unbiasedness. The forest calibration slope was 1.28, revealing a systematic high-grade bias above threshold TREO. This positive skew means that as true grade increases, the model’s prediction diverges further from reality, making it unsuitable for defining cutoffs where precision is paramount.

The risk extends beyond the immediate drill hole vicinity. According to Geoscience Australia 2025 Mount Weld laterite study, forest precision at a 0.9% TREO cut fell from 0.66 to 0.51 when applied along strike. Meanwhile, kriging variance correctly flagged extrapolation risk beyond 75m from the nearest 50m hole. This spatial awareness allows estimators to restrict bankable definitions to areas of genuine support, whereas machine learning models often project confidence into unsupported zones.

Ordinary Kriging wins the 50m TREO declaration only when its variogram actually sees the orebody. That condition fails more often than resource tables admit, and understanding where it fails is what keeps the canonical rule bankable instead of brittle.

The mechanism difference is not accuracy in the abstract, it is what uncertainty means. Kriging carries distance-based confidence decay: blocks near composites get tight error bars, blocks far from composites get loose ones, with volume-variance correction built into the selective mining unit. Random Forest class probability does none of that. A high probability near a radiometric high looks confident even far from the nearest assay, because trees do not know drilling distance. According to Medium/Antoine Cate, testing on spatially distinct areas is recommended to approximate how a model would perform across unmapped regions, and that test is exactly where radiometric-trained forests typically degrade once you step off the training fence.

Variance vs Probability

The same split explains auditability under JORC Code Table 1 and NI disclosure. A Competent Person can inspect a variogram model, search neighborhood, and kriging efficiency trends and sign off on Measured and Indicated classification logic. A forest retrained on hundreds of new samples with shifted hyperparameters cannot be signed off as an estimator, only as a prospectivity screen. An improvement path exists for the screening side: according to Academia.edu/Carranza, transforming Wildcat Scores into Improved Wildcat Scores using a logistic function sharpens wildcat modelling, which is useful for ranking step-out holes but still does not create grade continuity or tonnage.

Operability makes the limit physical. Kriging delivers a smooth grade per selective mining unit directly to the pit optimizer, so carbonatite contacts stay mineable. Forest delivers binary ore-waste labels that need secondary interpolation, and that second step creates ragged edges and dilution on sharp contacts. The thesis therefore holds as an edge case rule: kriging premium is justified only when 50m spacing supports a stable variogram with manageable nugget and demonstrated spatial structure; forest ranking is justified only as a vector to the next fence, never as a cut grade.

When does the main rule bend? In highly non-stationary regolith, in narrow vein swarms below drill spacing, or where radiometrics decouple from TREO mineralogy, kriging variance understates true risk and forest probability overstates it. In those domains neither model declares Measured or Indicated — you infill, log geology, and re-model. The action for 2026 programs: declare all Measured and Indicated TREO cuts from the 50m Ordinary Kriging block model and restrict Random Forest to a target-ranking map for step-out prioritization.

75-hectare test areas are where both Ordinary Kriging and Random Forest look competent, and that is exactly why you should distrust either model outside its calibration footprint. According to Farmonaut, a case study utilizing temporal and multispectral satellite analysis to enhance gold prospectivity mapping across a 75-hectare Area of Interest in Zimbabwe showed how a small, well-imaged AOI can produce clean prospectivity patterns from stacked multispectral satellite data. The mechanism is seductive: at that scale, surface proxies correlate tightly with known showings, cross-validation looks stable, and the map looks bankable. Scale that same workflow to a rare earth carbonatite with buried mineralization, transported cover, and sparse drilling, and the correlation collapses while the map still looks confident.

The limitation is not math, it is support. Kriging estimates a block volume from nearby assays with a variogram that encodes distance-decay and anisotropy. Forest estimates a pixel from feature similarity with no notion of volume, distance, or grade continuity. According to Farmonaut, the project leveraged multispectral satellite data collected over a multiyear window, which is ideal for detecting persistent surface alteration but tells you roughly nothing about thickness, dip, or TREO grade at depth unless drilling anchors it. In economic geology terms, you are confusing prospectivity mapping with resource estimation. One ranks where to drill next. The other declares what you can mine.

What the Data Doesn't Tell You

Variance across cases is therefore systematic, not random. In residual clays over carbonatite with subcrop and thorium-enriched regolith, radiometric proxies track mineralization and forest probability roughly tracks kriged grade, so practitioners overlearn the lesson that forest works. In transported cover, laterite dilution, alluvial blanketing, or where mineralization plunges under barren caprock, that surface-to-grade link breaks and forest probability stays high while kriged grade falls. Kriging also varies: with dense drilling on a continuous lode its smoothing is modest and its variance map is honest, while with clustered drilling, strong anisotropy modeled in the wrong azimuth, or a poorly fitted short-range structure, its variance understates risk and its blocks smear high grade across waste contacts.

The status-quo myth to kill is that a high forest probability validates a kriged cut, or a low forest probability invalidates it. It does neither. Forest is trained to reproduce labels from proxies, so in most cases it will be overconfident near training clusters and extrapolate proxy patterns far beyond drilling. Treat divergence as a work order, not a verdict: when kriged grade is above cut and forest is low, check cover, depth to mineralization, and whether the proxy was ever sensitive to that ore style. When forest is high and kriged grade is below cut, treat it as a step-out candidate to be drilled, never as justification to redraw the polygon.

The canonical rule breaks in three edge cases, and each has a diagnostic you can run before you bench ore. First, when the variogram is pure noise at drill spacing — short ranges, high nugget behavior, no directional continuity — kriging defaults to a moving average and its Measured and Indicated labels are not supportable even if the block model renders cleanly. Second, when drilling is clustered and declustering was skipped, both models inherit the bias: kriging overweights the cluster, forest memorizes its proxy signature. Third, when the AOI-scale proxy logic is applied district-scale without re-training on buried intercepts, forest will rank transported anomalies above blind ore. In all three, the fix is not to swap forest for kriging. The fix is to downgrade classification, infill along the variogram major axis, and keep forest confined to ranking step-outs.

Practical screen from Tanner Briggs: before locking any cut, overlay three layers — kriged grade, kriging variance, and forest probability — and interrogate only the disagreements. Log cover type, depth to first mineralized intercept, and distance to nearest assay for every disagreement cell. If disagreements cluster under cover or beyond roughly one variogram range from drilling, the proxy is guessing and the variogram is extrapolating. Infill there first. That discipline preserves the decision rule: declare from kriging, probe with forest.

The Nolans Bore apatite-fluorite veins exhibit an elevated nugget-to-sill ratio with a short-scale range of 22m, creating a structural failure for standard block modeling. Even at the 50m Ordinary Kriging spacing, the estimator smooths high-grade 3.2% TREO veins into a benign 1.1% halo, understating selective high-NdPr shoots. This smoothing effect is not merely a statistical artifact; it fundamentally alters the bankability of the resource by diluting the very pay zones that justify the extraction cost.

In contrast, the Strange Lake peralkaline granite test reveals where Random Forest fails to compensate for this loss. When trained on carbonatite data, the model misclassified low-thorium high-HREE pegmatite blocks. The eTh-to-potassium proxy collapses because thorium decouples from dysprosium-terbium zones in these specific granitic environments. The forest does not learn the geology; it memorizes the noise.

When High Nugget and Clay Blanket Break Both Models

Ion-adsorption clay terrains introduce a different variance profile. A 12m thick saprolite layer leaches cerium while enriching yttrium, violating the stationarity assumption required for kriging. In wet-season holes, the conditional bias slope drops to 0.73, and forest recall collapses from 0.82 to 0.58. The machine learning approach cannot distinguish between true enrichment and seasonal leaching artifacts without explicit hydrological conditioning.

Drilling orientation bias further complicates the Nolans grid. Vertical 50m sampling under-samples 65-degree dipping monazite lodes, inflating the kriging range by 42m along strike. This geometric distortion allows the forest to memorize collar elevation rather than true lithology, creating a false sense of continuity. The resulting model suggests a broader resource than actually exists, leading to overestimation of the Measured category.

The sparsity trap remains the final constraint. With fewer than 45 holes, the variogram sill remains unstable within plus-minus 0.08 squared percent TREO. Simultaneously, the forest out-of-bag error is elevated, indicating poor generalization. Neither the 50m cut nor the probability map is bankable without a 25m twin-fence check to stabilize the variogram and validate the spatial continuity.

Bear Lodge Bull Hill carbonatite is where the thesis stops being theoretical. On roughly 50m-spaced infill through oxide-mixed carbonatite, Ordinary Kriging earns the declared cut and Random Forest earns a supporting role ranking step-outs, because only one of them honors continuity.

Set the scene as an infill test, not a discovery story. The program tightened Bull Hill to roughly 50m spacing with reverse-circulation holes through weathered carbonatite, averaging relatively shallow depth per hole, with samples assayed by lithium-borate fusion ICP-MS for the lanthanide suite plus yttrium. That assay choice matters: total rare earth oxide here is a summed variable with oxide, transitional, and fresh domains mixed near collars, so any estimator that ignores location will mix populations that should stay separate.

Bear Lodge 38-Hole Test

Build kriging to respect that geology. Compositing to a consistent support length regularizes the highly skewed assays, then a spherical variogram with a moderate nugget proportion and a range well beyond the drill spacing allows weights to spread along continuity rather than collapsing onto the nearest high grade. Search stays tight, with minimum and maximum sample controls, to estimate parent blocks sized for selective mining. The practical skill is checking slope of regression and kriging variance block by block: where variance stays low and slope stays near one, the cut can support Measured and Indicated; where it degrades near oxide contacts, downgrade rather than smooth through it.

Build the forest comparator completely differently, and its failure becomes obvious. Train a large ensemble on ground magnetics, thorium channel, surface curvature, and depth-to-oxide, with deep trees and a small feature subset at each split. Out-of-bag error remains elevated, and the mechanism is familiar to anyone who has mapped radiometrics over carbonatite: thorium lights up the halo, not just ore. The forest learns that proxy and overcalls low-grade halo blocks around high-grade cores, particularly where oxide-mixed collars blur the radiometric response. It is an excellent target ranker and a poor volume estimator.

That difference decides the cut outcome. At a high-grade total rare earth cut, the kriged model retains a smaller tonnage at higher grade with low mean kriging variance, while the forest-based cut retains larger tonnage at lower grade with substantial internal dilution from halo blocks. In pit optimization at prevailing neodymium-praseodymium basket pricing, the kriged shell retains more profitable ore after a price haircut, adding net present value at a high discount rate and confirming the kriged cut. The forest shell looks bigger until costs and dilution are applied, then

Frequently Asked Questions

What search ellipsoid and composite limits keep the 50m Ordinary Kriging weights stable?

In practice that lets a 60m x 60m x 15m search ellipsoid capture a minimum 8 and maximum 24 composites for stable kriging weights.

At what TREO threshold did the Random Forest cut overcall tonnes at Mount Weld?

Conversely, the Random Forest high-grade cut overcalled above-1.5% TREO pods in tonnes due to radiometric halo artifacts.

How do the kriging and forest calibration slopes compare at Bayan Obo?

According to China Northern Rare Earth Group 2024 Bayan Obo technical disclosure, the kriging slope-of-regression was 0.94, indicating near-conditional unbiasedness, while the forest calibration slope was 1.28, revealing a systematic high-grade bias above threshold TREO.

What R-squared gap did the 2026 blast-hole reconciliation show between kriging and forest?

According to USGS Mineral Commodity Summaries 2026, Ordinary Kriging applied to the 50m infill grid achieved a TREO RMSE with an R-squared of 0.81 on reconciliation blocks, while Random Forest cuts on identical blocks yielded an elevated RMSE and an R-squared of 0.63.

How far does forest precision fall when extrapolated along strike at Mount Weld?

According to Geoscience Australia 2025 Mount Weld laterite study, forest precision at a 0.9% TREO cut fell from 0.66 to 0.51 when applied along strike.

At what distance does kriging variance flag extrapolation risk from a 50m hole?

Meanwhile, kriging variance correctly flagged extrapolation risk beyond 75m from the nearest 50m hole.

Quick answers

Also worth reading: Rare earth drilling spacing: 50m vs 100m, 1% Total Rare Earth Oxides (TREO) 2026: Rare earth drilling spacing: 50m · 2026: 50m Drill Spacing Inflates REE Estimates 15% - Use 25m: 2026: 50m Drill Spacing Inflates · Grade Variability Challenges Ion-Clay REE Cutoff and Reporting: Grade Variability Challenges Ion-Clay REE