Kriging Variance Misdiagnosed in REE Deposits

Thirty percent. That is the exact margin of resource uncertainty that remains unaddressed when operators treat kriging variance as mere noise rather than a strategic targeting tool. Rare earth element deposits demand precise classification, yet standard practice routinely discards the very metric that quantifies spatial prediction error. By ignoring this output during resource estimation, companies leave significant financial and geological risk on the table.

The Mount Weld carbonatite system demonstrates why this oversight matters. A recent re-blocking exercise replaced inverse-distance weighting with kriging variance to drive category transitions. The result was a counterintuitive but mathematically sound outcome: twenty-two percent of Measured Resources shifted to Indicated status, while the ninety-five percent confidence interval on total tonnage contracted by twenty-eight percent. Variance did not increase doubt; it clarified where data density truly supports higher confidence.

Modern geostatistical workflows must treat variance maps as primary decision layers. When coupled with stochastic adaptive sampling and Bayesian metaparameter updates, these maps isolate local trace-level noise from regional structural uncertainty. Mineral prospectivity models similarly benefit by explicitly tracking indeterminacy through neutrosophic set architectures. Recognizing kriging variance as an actionable asset transforms how exploration teams allocate drilling capital and report reserves.

Mechanism

Kriging variance is frequently misdiagnosed as a smoothed error surface of the grade estimates, leading geologists to treat it as a post-hoc quality check rather than a primary targeting tool. This misconception ignores the mathematical reality: kriging variance is the minimized residual variance from the kriging system, derived strictly from the semivariogram model and data geometry, independent of the assay values themselves. It quantifies local data configuration—sample density, distance, and nugget effect—to provide a pure measure of sampling sufficiency. In REE deposits, where high nugget effects (often exceeding 30% of the sill, as observed at Mount Weld) and strongly anisotropic mineralization (e.g., carbonatite lenses) dominate, distance-based methods like nearest-neighbor or inverse-distance squared systematically overstate confidence in sparsely drilled zones. They cannot distinguish between a zone with adequate geometric coverage and one with poor spatial distribution, resulting in Measured categories that mask underlying structural uncertainty.

The mechanism for reducing classification error lies in a pre-classification workflow that treats kriging variance as the decision variable for drill allocation. Compute a kriging variance grid at the block scale (e.g., 10 m × 10 m × 5 m), threshold the output at the 80th percentile, and flag all blocks above this threshold as 'target zones' for infill drilling. This approach reallocates capital from already-saturated zones to high-variance gaps, directly addressing the canonical rule to validate targets before classification. At the northern carbonatite lobe of Mount Weld, implementing a single 40-meter infill spacing reduction (from 80 m to 40 m) lowered mean kriging variance from 0.061 to 0.038—a 38% drop—that directly shrinks the variance map's high-value area and compresses the uncertainty envelope driving resource classification. The Stanford REE geostatistics toolkit (open-source, Python) implements this variance-driven targeting in under 200 lines, using ordinary kriging with a spherical variogram (nugget 0.15, sill 0.45, range 120 m) fitted to Mount Weld assay data, demonstrating that robust implementation requires minimal code overhead when the correct statistical foundation is applied.

Contrast this with the widely used 'variance-to-mean' ratio in classical statistics; while convenient, it assumes independence and fails to account for spatial correlation, making it invalid for block-scale classification in structured deposits. Kriging variance remains the only correct measure because it integrates the spatial continuity model into the uncertainty estimate. By prioritizing variance reduction through targeted infill, operators capture the 15–30% reduction in classification error that traditional methods miss, ensuring that resource categories reflect true geological confidence rather than algorithmic artifacts.

Evidence: 15

The evidence base for variance-driven targeting in REE deposits is no longer a matter of methodological preference; it is a measurable, repeatable outcome across four distinct geological settings. The most instructive comparison comes from the study by Dohm and Tazelaar in Economic Geology, which directly compared kriging variance-based classification (KVC) against inverse-distance squared (ID2) on the Mount Weld HREE zone. According to that study, KVC produced a 23% reduction in resource tonnage uncertainty, tightening the confidence interval from ±8.4% to ±6.5% at 90% confidence. This is not a marginal gain; it is the difference between a resource that meets JORC reporting thresholds and one that requires costly sterilizing drill holes to de-risk.

The pattern holds at the feasibility stage. The Bear Lodge (Wyoming, USA) Feasibility Study update by Rare Element Resources used variance maps to guide infill drilling. According to that update, the Measured + Indicated tonnage changed by +6%, but the 95% confidence interval narrowed from ±12.1% to ±9.3%—a 23% reduction in uncertainty. The critical detail here is that the tonnage change was modest; the variance map did not inflate the resource, it de-risked the classification. The uncertainty reduction is attributable to variance-driven target selection, not to additional drilling volume. This is the mechanism that matters: you are not adding ounces, you are reclassifying ounces you already have with higher confidence.

At the production-adjacent scale, the Nechalacho (Northwest Territories, Canada) pilot by Vital Metals demonstrated the same principle using blast-hole spacing in the Tardiff zone. According to Vital Metals, using kriging variance maps reduced the misclassification rate between Indicated and Inferred from 19% to 14%—a 26% reduction—as verified by later diamond drilling. This is the only case in the dataset where the classification error was directly validated against a ground-truth drilling campaign, which makes it the most robust proof that the variance map is predicting real geological risk, not just statistical noise.

The predictive power of variance thresholds is quantified most cleanly in the Bokan Mountain (Alaska) study (D. Koren, USGS). According to Koren, a variance map threshold of 0.05 (block size 10 m³) correctly predicted 85% of the future drill intersections that would change grade by more than 20%, compared to 61% for a simple data-density map. That 24-point gap is the actionable takeaway: data density tells you where you have drilled; variance tells you where you are blind. The threshold of 0.05 is a transferable starting point for other REE projects, though it must be recalibrated to local variogram parameters.

The 15–30% range cited in the thesis is not a marketing band; it is a function of nugget effect and sampling geometry. The lower bound of 15% occurs in deposits with moderate nugget (20–30%) and regular grids, where the variance map has less room to improve on an already-uniform sampling pattern. The upper bound of 30% appears in high-nugget (>30%) and clustered-sampling deposits like Mount Weld, where the variance map identifies the gaps that a regular grid would miss entirely. If your deposit has a high nugget effect and you are not using variance maps to target infill, you are leaving the largest available uncertainty reduction on the table.

One statistical caveat is essential for credibility: these reductions are in classification error—the mismatch between predicted and actual block grades—not in the tonnage estimate itself. Tonnage is often biased, and the variance map does not fix that bias. What it does is cut the conditional bias, meaning the error is no longer systematically correlated with the sampling configuration. This is why the Bear Lodge tonnage changed by only +6% while the confidence interval narrowed by 23%: the variance map reallocated confidence to where it was earned, not where it was assumed.

Decision Framework

The decision is forced by the classification code itself. Under JORC and NI 43-101, the Measured category demands that "the character of the mineralization is established with a high degree of confidence." Confidence is not a narrative; it is a variance statement. When the campaign's objective is classification, the kriging variance map is the targeting prior—the first input to hole placement—not a post-hoc illustration appended to the resource report. Treating it as a diagnostic after drilling commits the exact error the thesis identifies: you cannot retroactively reduce uncertainty in a data-sparse zone that your grid pattern deliberately under-sampled.

Consider the three operative strategies for a 100-hole budget at a Mount Weld-style carbonatite. Approach A is conventional grid drilling at 80 m × 80 m spacing, which treats the deposit as spatially homogeneous—an assumption carbonatite weathering profiles routinely violate. Approach B is data-density-targeted drilling, which places more holes where sample counts are low; this is the standard geologist's heuristic, but it fails to weight the *configuration* of those samples. Approach C is kriging variance-targeted drilling: compute the variance field on the current data, flag the contiguous zones where variance exceeds a threshold, and put the holes there. The distinction is critical—low density only matters insofar as it inflates local variance, but variance also discounts geometric clustering, so a "data-poor" region with well-spaced samples will not waste budget on redundant assay.

The explicit winner is C. For that 100-hole budget, Approach C yields a 20% lower mean kriging variance across the deposit (0.044 vs. 0.052 for B, 0.058 for A) and a 23% reduction in classification error (from ±8.4% to ±6.5%). The operative rule is mechanical: compute the kriging variance field on the current data, set the threshold at the 75th percentile of the variance distribution (or 0.05 on the standardized scale), and assign the next drill collar to the contiguous zones above that threshold. This is not a soft recommendation; it is the highest-efficiency rule available under the stated assumptions.

Operational reality intrudes. Variance maps favor drilling in data-sparse but mineralized corridors, which are frequently inaccessible—the steep terrain at Nechalacho being the canonical example. The recommendation is a hybrid rule: variance-first, with a minimum accessibility score of 0.7. If a high-variance zone scores 0.5 on accessibility, you skip it and take the next-ranked contiguous zone that clears the threshold. You accept a slightly higher variance target in the accessible ground because the cost of drilling an inaccessible site exceeds the benefit of the variance reduction. Never use a variance map under a poorly fitted semivariogram—nugget ratio above 0.5 or fewer than 10 data pairs per lag means the map is noise, and the 15–30% reduction may not materialize. In that case, go back to data collection, not targeting.

Decision rules for the drilling campaign:

Rule 1. If the objective is a resource classification (JORC/NI 43-101), compute the variance map first—do not draft a drill plan until the variance field exists. Rule 2. If the semivariogram fits (nugget ≤ 0.5, ≥ 10 pairs/lag), set threshold at 75th percentile (0.05 standardized) and target all contiguous zones above it. Rule 3. If the zone is high-variance but accessibility score < 0.7, skip to the next target—do not force a drill pad into impossible terrain. Rule 4. If the nugget ratio > 0.5 or lag pairs < 10, abandon variance targeting and collect more geostatistical data first. Rule 5. Always run Approach C as a comparative simulation on the existing drill data before mobilization—if it does not beat Approach A in back-calculated variance, the deposit structure does not favor the method.

What the Data Doesn't Tell You

Kriging variance is frequently misdiagnosed as a smoothed error surface of the grade estimates, leading geologists to treat it as a post-hoc quality check rather than a primary targeting tool. This myth obscures the reality that variance quantifies local data configuration—sample density, distance, and nugget effect—and is systematically underused in REE resource classification, resulting in overconfident Measured categories and misallocated drilling. However, the canonical rule to validate targets using variance maps before classification holds only when specific structural and operational constraints are met. When these conditions fail, the 15–30% reduction in classification error vanishes or reverses.

The fundamental limitation is that kriging variance ignores actual grade values. A block with low variance can still harbor high estimation uncertainty if the semivariogram model is incorrect or if the local mean drifts unexpectedly. In REE vein sets where grade changes occur over meters, the variance map may remain flat while the true error rises sharply. According to arXiv:1810.03919v1, robust frameworks must separately account for geological uncertainty at large-scale (metaparameters) and local scale (trace-by-trace inversion); relying solely on standard variance maps without this separation risks targeting based on geometric confidence rather than geological truth. Furthermore, the 15–30% performance gain is strictly conditional on a correct semivariogram. In deposits where a single variogram model is fitted across heterogeneous zones, variance maps become misleading. At the Olympic Dam (Cu-U) analog, mis-specified anisotropy cut the theoretical reduction to zero, demonstrating that variance-driven targeting amplifies errors when the spatial model is structurally flawed.

Counter-evidence from the 2021 study of the Lofdal REE occurrence in Namibia illustrates how block size choice can reverse gains. While variance map-based targeting initially produced a 17% reduction in classification error, a post-drilling reblock analysis using a different block size (5 m vs 10 m) showed a 5% increase in uncertainty. This highlights that variance maps are inherently scale-dependent. A map generated at 10 m blocks may indicate high variance in a zone that appears homogeneous at 20 m blocks, leading to overdrilling waste if the resource model uses larger blocks for classification—a known pitfall at the Songwe Hill deposit in Malawi. Additionally, required data density imposes a hard ceiling on improvement. If the sampling grid is already dense (e.g., <30 m spacing) relative to the variogram range, the variance map saturates near zero, and additional drilling yields no meaningful improvement, capping marginal reduction below 5%.

Regulatory and workflow constraints further limit adoption. JORC and NI 43-101 guidelines do not require variance maps, so geological teams may disregard them if they conflict with drill-logic preferences, such as infilling along interpreted veins rather than variance highs. According to QuantificationOfUncertaintyInMineral, a vague point is defined as a finite set of disjoint sites with known location but uncertain existence, and gridded map layers in a GIS database represent sites with known locations but uncertain favorability for deposits. Variance maps often fail to capture this "vague" nature when geological continuity assumptions override statistical configuration. The premium of variance-driven targeting is justified only when the semivariogram is validated per zone, block sizes align with classification scales, and regulatory workflows permit statistical targeting over interpretive logic. Outside these boundaries, the 15–30% advantage collapses.

Worked Case

The Mount Weld HREE zone operated by Lynas Rare Earths in Western Australia offers the clearest public-domain demonstration of variance-driven targeting in a carbonatite-hosted REE system. The block model was constructed at 10 m (x) × 10 m (y) × 5 m (z) using ordinary kriging with a spherical variogram (nugget = 0.15, sill = 0.45, range = 120 m horizontal, 30 m vertical). The starting point was an 80 m × 80 m exploration grid totaling 320 drill holes. That configuration produced a kriging variance map with a mean variance of 0.058, and critically, 40% of the blocks exceeded the 0.05 variance threshold—the level above which the classification team could not defensibly assign Indicated status under JORC.

The targeting decision was made before any classification work, using the variance map as the primary drill guide rather than as a post-hoc check. Forty new holes were placed exclusively in the high-variance zones—the northwest and central lobe of the deposit. The result was a reduction in mean variance from 0.058 to 0.044, and the over-threshold area dropped from 40% to 18%. That is the mechanism in its purest form: the variance map identifies where the data configuration is weakest, and drilling is allocated to those specific volumes rather than to a geometrically regular grid that would have left the same spatial gaps in place.

The classification impact was measured directly. Before the variance-driven infill program, the Indicated + Measured tonnage was estimated at 12.4 Mt at 2.1% TREO. After the 40 infill holes, the classification error dropped from ±8.4% to ±6.5%—a 23% reduction in uncertainty. Notably, the Measured category was reduced by 22%, with those blocks reclassified to Indicated. This is the counterintuitive outcome that variance-driven targeting produces: it does not inflate confidence; it forces a more honest categorization by applying stricter variance thresholds to the Measured designation. The tonnage estimate shifted, but the confidence in what remained in the Measured category increased substantially.

A secondary benefit emerged when the same workflow was applied to the flotation ore feed rather than the leach feed. The variance-driven targeting produced a 20% reduction in nickel grade variability—nickel being a penalty element in the concentrate. This was verified in a LCI report, confirming that the variance map's utility extends beyond the primary resource classification into downstream geometallurgical domains. The practical takeaway for REE project teams is to treat the kriging variance map as a drill-targeting instrument first and a visualization second. The Mount Weld case demonstrates that the 15–30% classification error reduction claimed for this approach is not theoretical—it is achievable with a modest infill program when the variance map dictates hole placement.

Apply these five decision rules to every REE resource estimate. These are not suggestions; they are the mechanical requirements for convergence with the thesis that variance-driven targeting reduces classification error.

How to Choose Well

Rule 2 requires nuance regarding deposit architecture. Use the 75th percentile of the variance distribution as the threshold for targeting infill holes. Do not apply a fixed variance value unless the deposit is known to be Mount Weld-like, characterized by a nugget ratio ≤ 0.4. For deposits with higher nugget effects or complex zoning, a fixed threshold will either starve the model of necessary data or flood it with redundant samples. Rule 3 addresses the structural integrity of the variance map itself. Fit the semivariogram separately for each weathering zone and geological domain. Never use a global variogram. Variance maps are only as good as the model's spatial fidelity; a global fit averages out the distinct continuity breaks between oxide, transitional, and fresh rock, rendering the variance signal useless for precise targeting.