Resource estimation is the most consequential analytical workflow in mining: the resource estimate is the basis for project valuation, financing, and ultimately the mine plan. Methodology in this space changes slowly, for good reasons. The classical geostatistical framework — variography, kriging, conditional simulation — has decades of validated use, well-established professional standards, and a Qualified Person community that understands its assumptions, limitations, and edge cases. Machine learning methods coming into the workflow have to clear a high bar of demonstrated reliability before they replace established practice.
The honest summary of where things stand in 2026 is this: ML methods have earned a real role in certain parts of the resource estimation workflow, particularly in geological domaining, data validation, and exploratory analysis. They have not displaced classical geostatistics in the estimation step itself, and the published evidence increasingly suggests that they shouldn't. The future of resource estimation is hybrid — classical methods where they're proven, ML methods where they add demonstrable value, with the QP making explicit choices about which methodology applies where.
Why Geostatistics Has Held Up So Well
The reason classical geostatistics has resisted displacement is not institutional conservatism — though that's part of it — it's that the underlying methods are well matched to the problem. Kriging is a best linear unbiased estimator under a clearly stated set of assumptions, and the standard variography workflow makes those assumptions explicit and inspectable. The methods produce not just an estimate but a quantified uncertainty in the form of kriging variance or conditional simulation realizations. The output is directly usable for resource classification, mine planning, and risk analysis in ways that ML black-box predictions are not.
The methodology also has decades of QP experience attached to it. The professional community understands when kriging works, when it breaks down (extreme grades, sparse data, complex anisotropy), and how to adjust the workflow for these cases. The professional standards that govern resource classification — Measured, Indicated, Inferred — are operationally defined in terms that map naturally onto geostatistical outputs (drill spacing, kriging variance, conditional simulation confidence). Replacing the underlying methodology means rebuilding this professional framework, which is a higher bar than any individual technique can usually clear.
And finally, the failure modes of classical geostatistics are well documented. When kriging produces a smoothed estimate that misses high-grade pockets, the QP knows to apply post-processing or conditional simulation. When the data fails the stationarity assumption, the QP knows to compartmentalize into separate domains. ML methods have failure modes too, but they're less catalogued and harder to predict in any given project context.
Where ML Has Genuinely Earned a Role
Three places in the resource estimation workflow where ML methods are now doing real work in published industry case studies. First, geological domaining — the process of subdividing a deposit into compositionally distinct sub-volumes before estimating each separately. Domain definition has always been part data analysis, part geological interpretation. ML clustering methods — applied to multi-element geochemistry, lithology codes, alteration intensity, and structural data from drill logs — can propose domain boundaries that integrate more variables than visual interpretation would and surface domain distinctions that aren't obvious from any single variable. The QP reviews and adjusts; the ML does the heavy lifting of multi-variable pattern recognition.
Second, data validation and quality control. Before any estimation begins, the QP has to be confident in the data: assays correctly assigned to drill holes, no transcription errors, no analytical biases, no compositing artifacts. ML-based anomaly detection methods catch outliers and inconsistencies that manual review misses, and they do it consistently across large datasets. The pattern is the same as for QA/QC of fresh geochemistry: the methods don't replace the QP's review, they make the review more thorough.
Third, in non-stationary and complex deposit geometries where classical variography is difficult. ML methods, particularly those that learn spatial structure rather than assuming a single variogram model, can produce estimates that respect complex geological controls better than naive kriging. This is most valuable in deposit types — vein networks, structurally controlled mineralization, complex polyphase systems — where the assumptions of stationary geostatistics are weakly satisfied. The ML methods are not necessarily better than careful classical geostatistics with appropriate domaining; they're more forgiving when the analyst has limited time to do the domaining work properly.
Where ML in the Estimation Step Itself Stumbles
The case for ML in the estimation step — the actual production of grade estimates at the model block level — is weaker than its promotional materials would suggest. Several specific problems show up repeatedly.
Most ML methods do not naturally produce calibrated uncertainty. A random forest's prediction at a block is a point estimate; the per-tree variation can be turned into an uncertainty estimate, but the result is not a probability distribution in the way conditional simulation output is. For resource classification, which is fundamentally about uncertainty, this is a serious limitation. Workarounds exist — quantile regression forests, Bayesian neural networks, post-hoc calibration — but they're additional methodology that needs its own validation.
ML methods can overfit local data in ways that classical geostatistics doesn't. A random forest trained on closely spaced drill data will reproduce the training data essentially perfectly while extrapolating poorly to sparsely sampled blocks. Cross-validation can detect this, but it doesn't necessarily fix it without methodology that's structurally different from off-the-shelf ML.
Smoothing behavior in ML methods is hard to control and easy to misread. Kriging's smoothing is a well-understood property with documented impact on resource classification; ML smoothing varies by method, by hyperparameter setting, and by data characteristics in ways that are harder to predict. A resource estimate that's been smoothed in unknown ways is a poor basis for downstream planning.
And the regulatory and professional acceptance of ML-only resource estimates remains uncertain. NI 43-101 and CIM Best Practice Guidelines are written in terms that map naturally onto classical geostatistics; ML estimation requires the QP to demonstrate equivalence to the standard framework, which is additional defensibility work that doesn't apply when classical methods are used.
The Recent Published Evidence
The academic and industry literature on ML in resource estimation has grown substantially in the past several years, and the conclusions are converging. Comparative studies — where the same dataset is estimated with both classical geostatistics and ML methods, and the results compared against post-mining reconciliation data — generally find that the best-tuned ML methods perform comparably to careful classical kriging, while underperforming when applied naively without domain knowledge.
The studies that show ML clearly outperforming classical methods tend to compare ML to poorly-tuned or naive geostatistics; the studies that show classical methods clearly outperforming ML tend to compare to off-the-shelf ML applied without domain awareness. The honest read of the literature is that methodology quality matters more than methodology choice. A poorly-done study is poorly-done regardless of which methodology category it's in.
For QPs, the practical implication is that ML methods can be incorporated into resource estimation when the QP has time to validate them carefully against classical benchmarks, but the marginal value of ML over careful classical work is often small. The methodology change is worth making when there's a specific reason — complex geology that classical methods handle poorly, large datasets where ML scales better, or specific workflow integration benefits — not by default.
Implicit Modeling and the Geology Side
Adjacent to the estimation question, ML has had clearer impact on implicit geological modeling — the modeling of geological surfaces and volumes that precedes resource estimation. Leapfrog Geo, Geoscience ANALYST, and similar implicit modeling platforms use radial basis functions and related methods that share some mathematical structure with kernel-based ML. The progression from explicit (manually drawn) to implicit (mathematically interpolated) modeling has been a substantial methodological improvement over the past two decades, and the ongoing integration of ML methods into the implicit modeling step is producing modeled volumes that integrate more data more consistently than manual modeling could.
This benefits resource estimation indirectly: better geological models produce better domain definitions, which produce better-constrained resource estimates regardless of what method the estimation itself uses. The leverage of ML in this part of the workflow is meaningful and is being adopted faster than ML in the estimation step itself.
What QPs Are Actually Saying
The CIM and the professional engineering bodies have been publishing increasing guidance and discussion papers on the use of ML in resource estimation. The consistent message: ML is a tool the QP can choose to use, subject to the same professional standards of defensibility that apply to any methodology choice. The QP must understand the method, validate it appropriately for the dataset, and be able to explain why the chosen methodology is appropriate.
The QP's documentation burden when using ML is higher than when using established classical methods, because the ML methodology is less standardized and requires more justification. This is one of the practical reasons ML hasn't displaced classical methods faster: the marginal documentation cost often exceeds the marginal analytical benefit, and QPs default to the methodology that requires less justification.
A Reasonable Position
For a junior with a maiden resource on the way, the recommended posture is: classical geostatistics for the estimation step, ML methods as supplements for domaining, validation, and uncertainty analysis. Hire a QP who is comfortable with both methodology families and can make explicit choices about which applies where. Document the methodology choices in the technical report with the same rigor that classical-only methods would receive.
For a major with a long-cycle resource program, the right pattern is more experimental: pilot ML methods on selected deposits where the team has time to validate carefully, build internal capability to use them defensibly, and integrate them into the workflow where the validation supports it. The cost of building this capability is meaningful but it positions the company well as the regulatory and professional acceptance of ML methods continues to evolve.
For consulting support on integrating ML appropriately into resource estimation workflows without compromising QP defensibility, our free workflow audit covers technical workflows, or contact us for a deeper conversation.