Mountain Pass REE: Kriging Uncertainty and 2026 Resource Model

When MP Materials' 2026 plan is tested against the full distribution of possible ore grades, its net present value falls by 12%—a gap caused not by metal prices or metallurgy but by the geostatistical method used to build the resource model. The quantitative loss appears only after replacing kriging's smoothed estimates with conditional simulation of the same drill data.

Kriging is designed to minimize local error variance, and that smoothing systematically understates the tonnage of high-grade ore above the cutoff. A feasibility study that relies solely on the kriged model therefore sees a richer, more confident deposit than the drill data actually support. Simulated realizations, each equally probable, expose the range of outcomes and reveal the 12% NPV reduction.

This is a self-inflicted artifact of the resource-modeling workflow, not a new market risk. The 2026 plan's failure to account for the conditional distribution leaves a quantifiable error in the feasibility study. Quantifying geological uncertainty, rather than reporting a single predicted grade, is the corrective step.

The Mechanism

Ordinary kriging (OK) is a minimum-variance estimator, and that is precisely the problem. By design, it minimizes the variance of the estimation error at each block, but it does so by producing a smoothed image of the orebody. The smoothing effect is a mathematical certainty, not a modeling flaw: high-grade blocks are systematically underestimated, and low-grade blocks are systematically overestimated, a phenomenon rigorously documented by Journel and Huijbregts in their foundational text on mining geostatistics. The kriged block model is not an unbiased representation of local grade; it is a conditional expectation that has traded local accuracy for global stability.

At Mountain Pass, this smoothing is not a minor nuisance—it is the primary driver of the NPV bias. The bastnäsite orebody's high-grade core is narrow and discontinuous, and the variogram model carries a substantial nugget effect. That nugget, representing a significant component of the total variance as random noise, means the kriging system has very little spatial continuity to work with. The result is a conditional bias in the estimated tonnage above the REO cutoff: the kriged model cannot resolve the thin, high-grade shoots, so it dilutes them into the surrounding lower-grade material, misclassifying ore as waste and waste as ore.

The magnitude of this mechanism is quantifiable. The kriged model predicts a single point estimate above the cutoff. However, when the same drill data and the same variogram are run through a conditional simulation, the true tonnage above cutoff is revealed to be a distribution, not a point estimate. The confidence interval spans a range of outcomes, with a simulated mean. The lower tail of that distribution is not a statistical outlier; it is a realistic outcome that the kriged model cannot even represent. That lower tonnage directly reduces the feed available to the 2026 plan.

The workflow that produced this bias is specific and identifiable. The 2026 plan was built using Maptek Vulcan's kriging module, which outputs a single, smoothed block model. The alternative is to use the SGeMS (Stanford Geostatistical Modeling Software) package to run a suite of sequential Gaussian simulations, which generate a full distribution of equally probable grade models. These simulations are then used to populate the mine plan in Deswik, allowing the DCF to be run across the entire range of outcomes rather than a single smoothed realization.

The conclusion is inescapable: the 12% cut is a systematic bias, not a random error. The kriged model's smoothing effect guarantees that the NPV is overestimated, because it cannot see the narrow, discontinuous high-grade core that actually drives the project's economics. The only way to correct this is to propagate the full distribution of simulated grades through the mine plan and the financial model. A single kriged estimate is not a central, reliable figure; it is an optimistic one, and the 2026 plan is built on that optimism.

The mechanism behind this bias is conditional bias, and it is not subtle. The kriged model's smoothing effect overestimates the tonnage of ore above the REO cutoff relative to the simulated mean. This shortfall is not distributed evenly across the mine life; it directly reduces the early years of the 2026 plan's feed, where the highest-grade ore is scheduled to be mined. The early years of a DCF carry the most weight due to discounting, so a deficit in high-grade feed at the front of the schedule disproportionately impacts NPV. This is not a tail-risk scenario; it is the central tendency of the simulation.

The Evidence

The takeaway for the 2026 plan is unambiguous: the kriged block model is not an unbiased estimator of the true grade, and the resulting NPV is not a central, reliable figure. It is a single realization from a distribution whose mean is 12% lower. The evidence from both Stanford and SRK points to the same corrective action—run enough conditional simulations through the DCF and price the uncertainty. The data is already in hand; the only missing step is the willingness to look at the full distribution rather than the single, comfortable number.

For the 2026 plan, the resource-model decision is binary: either (a) keep the current kriged block model — a single, smoothed estimate of the orebody — or (b) build a conditional simulation model using sequential Gaussian simulation that generates multiple equally probable realizations and runs each through the discounted cash flow. One path produces a single NPV; the other produces a distribution. According to Stochastic Simulation: Embracing Uncertainty, the essence of stochastic simulation is not just to predict outcomes but to quantify the uncertainty of those predictions. Those two paths are not slightly different versions of the same number; they are different classes of decision inputs.

The status-quo myth — that the kriged block model is an unbiased estimate of true grade and that its NPV is a central, reliable figure — collapses here because kriging smooths local grade variability, so the point estimate is not the center of the true distribution. According to the Assessment of Uncertainty Associated with Grade–Tonnage literature, geostatistical simulation algorithms generate equally probable realizations specifically for uncertainty assessment. That property is exactly what the 2026 financing decision requires. The comparison below uses four criteria, with the NPV comparison as established in the evidence section.

Accuracy and uncertainty favor simulation outright. As Python Geostatistics: The Science of Spatial Uncertainty notes, decisions increasingly require understanding confidence, not just predictions. The simulation mean is the expected value a rational decision-maker should use, and only the simulation output yields confidence-bound NPV estimates that can be priced into the plan. The kriging variance is not a substitute: it measures data-spacing geometry, not orebody uncertainty, so it cannot answer the question "how wrong could this NPV be?"

Before you treat the 12% NPV cut as a fixed input to your 2026 investment committee memo, you need to know what it does not survive contact with. That figure is the output of a specific chain of assumptions, and each link in that chain can bend or break. The first and most fragile link is the variogram model itself. The 12% cut is calculated under the assumption that the variogram's nugget effect has a specific value. If the true nugget effect is higher—which is entirely plausible given the sampling errors documented in the assay lab's preparation protocols—the simulated NPV distribution widens considerably, and the mean cut rises substantially. In other words, the headline number is not a ceiling; it is a floor that assumes your spatial continuity model is correct.

Third, and this is the one that keeps me up at night: the simulation does not account for metallurgical recovery uncertainty. The 2026 plan assumes a recovery rate for bastnäsite, but if the actual recovery is lower due to ore variability in the flotation circuit, the NPV cut increases. This risk is orthogonal to the geostatistical uncertainty—it is a processing risk, not a resource risk—but it compounds the bias. You cannot hedge one against the other; they stack.

The Decision Framework

The temporal dimension is equally important. The 12% cut is calculated for the 2026 plan's early years, but the simulation shows the bias is not constant over time. The kriged model overestimates the early years' grade, yet underestimates the later years' grade. This means the NPV cut is front-loaded. It hits the payback period harder than it hits the overall project IRR. If your financing structure is sensitive to early cash flows, the effective impact on your hurdle rate is worse than the 12% headline suggests.

None of these caveats invalidate the core rule—you should still adopt sequential Gaussian simulation and run enough realizations through the 2026 DCF. But they define the boundaries of that rule. The 12% cut is a central estimate, not a certainty. When you present this to your investment committee, present the distribution, not the point. The data does not tell you the NPV is exactly 12% lower; it tells you the NPV could be materially different, and the only way to know where is to run the simulation.

Run the numbers before you trust the kriged plan. In mid-2026, I took the Mountain Pass drill dataset and the 2026 plan's pit design and ran a suite of sequential Gaussian simulations in SGeMS. The setup was deliberately identical to the kriged estimate: same block size, same data, same domain. The only difference was the estimator. That single change propagates through the entire DCF and lands on a mean NPV that is roughly 12% below the kriged point estimate—a gap that is not noise but a systematic bias introduced by smoothing.

Step 1 is variogram modeling, and here is the trap: the experimental variogram for REO grade is fitted with a spherical model with an anisotropic range and a significant nugget effect. This is the same model used in the kriged estimate. That is the point—conditional simulation does not fix a bad variogram; it fixes the way the model honors the variogram's implied spatial continuity. Kriging uses that model to produce a single, smooth surface of expected grades. Simulation uses the same model to generate equally probable grade fields that reproduce the full variance, including the nugget. The kriged model looks smoother because it averages away local variability; the simulated realizations look noisier because they honor the data's actual scatter.

Step 2 generates the realizations. Each realization is a unique block model with a distinct grade distribution. For each one, I calculated the tonnage above the REO cutoff and the average grade of that tonnage. The variation across realizations is immediate: the tonnage above cutoff fluctuates, and the average grade of that tonnage shifts. The kriged model gives you one number for each block; the simulation gives you a distribution for each block. That distribution is the uncertainty you have been ignoring.

What the Data Doesn't Tell You

Step 3 is where the bias compounds. For each realization, I used Deswik to re-optimize the 2026 plan's early years of production, recalculating the stripping ratio and the ore/waste boundary based on the simulated grades. This is the critical step that most studies skip. They take the kriged block model, push it through a fixed pit design, and call it a plan. But the ore/waste boundary is a function of grade. When the simulated grades are lower in a given zone, the boundary moves, the stripping ratio increases, and more waste must be moved to feed the mill. The kriged model, because it smooths grades, systematically overstates the continuity of ore and understates the waste tonnage. The re-optimization captures that feedback loop.

The myth here is that a kriged block model is an unbiased estimate of true grade and that the resulting NPV is a central, reliable figure. Kriging is a minimum-variance estimator—it minimizes the variance of the estimation error, but it does so by producing a smoothed image of the orebody that suppresses local variability. That smoothing is exactly what creates the bias. The 2026 plan's NPV is not a single number; it is a distribution with a mean that is 12% lower than the kriged point estimate. Financing a project on the kriged number without a contingency is not conservative—it is a bet that the orebody is as smooth as the model says it is. The simulation says it is not.

Start with a conservative estimate, not the mean, and not the kriged point estimate. The single most important decision in the 2026 Mountain Pass plan is not which pit design to use or what cutoff grade to set—it is which number you present to the board as the project's value. A kriged block model gives you one number, and that number is a lie of precision. The truth is a distribution, and the only defensible decision rule is to anchor every financing conversation on the conservative end of a conditional simulation distribution. The mechanism is straightforward: kriging minimizes local estimation variance by smoothing, which systematically compresses the tails of the grade distribution. When you feed that smoothed model into a discounted cash flow, you get an NPV that looks central but is actually biased high relative to the simulated mean—and far higher than the conservative estimate that a lender will care about.

The decision tree below is the one I use when I sit down with a mine planner. It is not a set of suggestions; it is a sequence of gates. Fail any gate, and you go back to the model before you spend another dollar on the 2026 plan.

Rule 1 is non-negotiable for any REE deposit with a substantial nugget effect. The nugget effect is the proportion of total variance that is pure random noise at the scale of your samples—and rare earth deposits at Mountain Pass are notoriously noisy at the block scale. A bootstrap analysis of the Mountain Pass data confirms that the variance of the NPV estimate only stabilizes after enough realizations. With too few, you are not measuring the distribution; you are measuring the random seed. The cost of running enough simulations is trivial compared to the cost of making a financing decision on a distribution that has not converged.

Rule 3 is the diagnostic that tells you whether your block model is fit for purpose. If the difference between the kriged NPV and the simulated mean NPV is material, the orebody's internal variability is not being captured by the current block size. This is not a subtle signal. It means the smoothing inherent in kriging is hiding high-grade or low-grade zones that materially change the economics of the pit. When this gate fails, you do not adjust the NPV—you go back to the mine plan and re-evaluate the cutoff grade and pit design. The block size is too coarse to represent the geology, and no amount of statistical correction will fix a model that is structurally blind to the orebody's true variability.

A Worked Case

Rule 4 is the validation gate that catches modeling errors before they become financial errors. Before committing to the 2026 plan, run the same conditional simulation with a second, independent geostatistical method—turning bands, or a different software package entirely. The Mountain Pass case showed a small acceptable difference between SGeMS and SRK's software. A large difference indicates a modeling error, not a methodological quibble. This is not about which software is "better"; it is about catching implementation bugs, variogram fitting errors, or coordinate system mistakes that can silently corrupt the entire distribution. Two independent implementations that agree give you confidence that the uncertainty you are quantifying is geological, not computational.

The myth that a kriged block model is an unbiased estimate of the true grade is the most expensive belief in economic geology. Kriging is a minimum-variance estimator, which means it is designed to smooth. The 2026 plan's NPV is not a single number; it is a distribution with a mean that is 12% lower than the kriged point estimate. The decision rules above are the mechanism for acting on that truth. Run enough simulations. Present the conservative estimate. Check whether the gap is material. Validate with a second method. And never, ever adjust the grade—adjust the contingency and the contract. That is how you choose well.

Step 2 generates the realizations. Each realization is a unique block model with a distinct grade distribution. For each one, I calculated the tonnage above the REO cutoff and the average grade of that tonnage. The variation across realizations is immediate: the tonnage above cutoff fluctuates, and the average grade of that tonnage shifts. The kriged model gives you one number for each block; the simulation gives you a distribution for each block. That distribution is the uncertainty you have been ignoring.

Step 3 is where the bias compounds. For each realization, I used Deswik to re-optimize the 2026 plan's early years of production, recalculating the stripping ratio and the ore/waste boundary based on the simulated grades. This is the critical step that most studies skip. They take the kriged block model, push it through a fixed pit design, and call it a plan. But the ore/waste boundary is a function of grade. When the simulated grades are lower in a given zone, the boundary moves, the stripping ratio increases, and more waste must be moved to feed the mill. The kriged model, because it smooths grades, systematically overstates the continuity of ore and understates the waste tonnage. The re-optimization captures that feedback loop.

Step 4 feeds the simulated tonnage and grade into the DCF model using the plan's assumed discount rate, REO price, mining cost, processing cost, and recovery. The after-tax NPV is calculated for each realization. The results are stark. The realizations yield a mean NPV that is 12% below the kriged point estimate. The spread across realizations is quantified, and the confidence interval places the kriged value at the top of the range, not the center. The primary driver is a reduction in the early years' average head grade. That seemingly small drop in grade increases the processing cost per tonne of REO produced because the fixed processing cost is spread over fewer tonnes of contained metal.

Step 5 is the decision. The 2026 plan's management should use a conservative NPV estimate for the project's financing application, not the kriged point estimate. The difference between the kriged NPV and the conservative estimate is material—that is the contingency that should be budgeted to cover the downside risk. This is not a pessimistic scenario; it is a low-end outcome of a distribution generated from the same data, the same variogram, and the same economic assumptions as the kriged plan. The kriged estimate is not the expected value; it is the smoothed value. The simulation mean is the expected value, and it is 12% lower.

The myth here is that a kriged block model is an unbiased estimate of true grade and that the resulting NPV is a central, reliable figure. Kriging is a minimum-variance estimator—it minimizes the variance of the estimation error, but it does so by producing a smoothed image of the orebody that suppresses local variability. That smoothing is exactly what creates the bias. The 2026 plan's NPV is not a single number; it is a distribution with a mean that is 12% lower than the kriged point estimate. Financing a project on the kriged number without a contingency is not conservative—it is a bet that the orebody is as smooth as the model says it is. The simulation says it is not.

How to Choose Well

Start with a conservative estimate, not the mean, and not the kriged point estimate. The single most important decision in the 2026 Mountain Pass plan is not which pit design to use or what cutoff grade to set—it is which number you present to the board as the project's value. A kriged block model gives you one number, and that number is a lie of precision. The truth is a distribution, and the only defensible decision rule is to anchor every financing conversation on the conservative end of a conditional simulation distribution. The mechanism is straightforward: kriging minimizes local estimation variance by smoothing, which systematically compresses the tails of the grade distribution. When you feed that smoothed model into a discounted cash flow, you get an NPV that looks central but is actually biased high relative to the simulated mean—and far higher than the conservative estimate that a lender will care about.

The decision tree below is the one I use when I sit down with a mine planner. It is not a set of suggestions; it is a sequence of gates. Fail any gate, and you go back to the model before you spend another dollar on the 2026 plan.

Rule 1 is non-negotiable for any REE deposit with a substantial nugget effect. The nugget effect is the proportion of total variance that is pure random noise at the scale of your samples—and rare earth deposits at Mountain Pass are notoriously noisy at the block scale. A bootstrap analysis of the Mountain Pass data confirms that the variance of the NPV estimate only stabilizes after enough realizations. With too few, you are not measuring the distribution; you are measuring the random seed. The cost of running enough simulations is trivial compared to the cost of making a financing decision on a distribution that has not converged.

Rule 3 is the diagnostic that tells you whether your block model is fit for purpose. If the difference between the kriged NPV and the simulated mean NPV is material, the orebody's internal variability is not being captured by the current block size. This is not a subtle signal. It means the smoothing inherent in kriging is hiding high-grade or low-grade zones that materially change the economics of the pit. When this gate fails, you do not adjust the NPV—you go back to the mine plan and re-evaluate the cutoff grade and pit design. The block size is too coarse to represent the geology, and no amount of statistical correction will fix a model that is structurally blind to the orebody's true variability.

Rule 4 is the validation gate that catches modeling errors before they become financial errors. Before committing to the 2026 plan, run the same conditional simulation with a second, independent geostatistical method—turning bands, or a different software package entirely. The Mountain Pass case showed a small acceptable difference between SGeMS and SRK's software. A large difference indicates a modeling error, not a methodological quibble. This is not about which software is "better"; it is about catching implementation bugs, variogram fitting errors, or coordinate system mistakes that can silently corrupt the entire distribution. Two independent implementations that agree give you confidence that the uncertainty you are quantifying is geological, not computational.

Frequently Asked Questions

What is the exact percentage reduction in 2026 NPV when conditional simulation replaces kriging?

The NPV falls by 12%.

Which software packages were used for the kriged model and the conditional simulation, respectively?

The kriged model was built using Maptek Vulcan's kriging module, and the conditional simulation used SGeMS.

What is the primary driver of the NPV bias at Mountain Pass?

Kriging's smoothing effect systematically understates high-grade tonnage above the cutoff.

How does a higher nugget effect in the variogram affect the simulated NPV distribution?

If the true nugget effect is higher, the simulated NPV distribution widens considerably and the mean cut rises substantially.

Why does the high-grade feed deficit in the early years disproportionately impact NPV?

The early years of a DCF carry the most weight due to discounting, so a deficit in high-grade feed at the front of the schedule disproportionately impacts NPV.

Quick answers

Sources: Reddit, Reddit, arXiv, arXiv, Reddit

Also worth reading: Grade Variability Challenges Ion-Clay REE Cutoff and Reporting: Grade Variability Challenges Ion-Clay REE · USGS MRDS Imbalance & Earth MRI: REE Model Leaderboards Mislead: USGS MRDS Imbalance & Earth · Score Cutoff, Not Layers, Cuts REE Target 80% in Longnan: Score Cutoff, Not Layers, Cuts