Show notes & full bibliography · Transcript

Both hosts are synthetic voices and the script is AI-generated. Every paper cited was verified against a primary source (DOI or arXiv ID) before publication, but the commentary is not that of practising researchers.

The Top-Cut — Show Notes

AI Meets the Block Model: Machine Learning in Resource & Reserve Estimation (Sep 2025 – Sep 2026)

Two hosts, Traci and Joao, go through roughly twenty verified papers from the

last twelve months at the intersection of mineral/ore resource and reserve

estimation (grade estimation, block modeling, geostatistics, conditional

simulation) and AI/ML methods applied to that specific problem — not

exploration targeting. Every citation below was checked against a primary

source (DOI, arXiv ID, or publisher page) before this episode was finalized;

where a claim in the audio was corrected live on air, that correction is

reflected here too, not the original (wrong) version.

Bibliography

Foundational works

These predate the review window but the episode builds directly on them, so

they are cited properly rather than name-dropped.

• Krige, D.G. (1951) "A statistical approach to some basic mine valuation problems on the Witwatersrand." Journal of the Chemical, Metallurgical and Mining Society of South Africa 52(6):119–130. The empirical origin of kriging: block-grade estimates must be regressed toward the mean to correct conditional bias. (No DOI for the original; a digitised record exists via journals.co.za.)
• Matheron, G. (1963) "Principles of geostatistics." Economic Geology 58(8):1246–1266. DOI: 10.2113/gsecongeo.58.8.1246. Formalised regionalized variables — variogram, estimation variance, optimal linear estimator — and named the method after Krige.
• Cheng, Q., Agterberg, F.P., Ballantyne, S.B. (1994) "The separation of geochemical anomalies from background by fractal methods." Journal of Geochemical Exploration 51(2):109–130. DOI: 10.1016/0375-6742(94)90013-290013-2). The concentration–area (C–A) fractal method.
• Chen, W., Li, Y., Reich, B.J., Sun, Y. (2020/2024) "DeepKriging: Spatially Dependent Deep Neural Networks for Spatial Prediction." arXiv:2007.11972; Statistica Sinica 34(1):291–311. DOI: 10.5705/ss.202021.0277. The original DeepKriging — the 2025 Erten & Boisvert paper below extends this, it did not originate it.

The 2018–2023 wave — geology as a feature

The episode's central distinction. These papers did not ignore geology; they

used it as an input variable rather than as a hard estimation constraint —

and consistently found it was where the accuracy came from.

• Jafrasteh, B., Fathianpour, N., Suárez, A. (2018) "Comparison of machine learning methods for copper ore grade estimation." Computational Geosciences 22(5):1371–1388. DOI: 10.1007/s10596-018-9758-0. RF, NN, Gaussian processes vs. kriging at Sarcheshmeh porphyry Cu. Coordinates-only baseline; reports significant gains when rock type is added as a predictor.
• Kaplan, U.E., Topal, E. (2020) "A New Ore Grade Estimation Using Combine Machine Learning Algorithms." Minerals 10(10):847. DOI: 10.3390/min10100847. kNN predicts lithology/alteration, then a neural net predicts grade from coordinates + those labels. Coordinates-only R²=0.112 → with geological labels R²=0.528 (MAE 0.862 → 0.507).
• Zhang, S.E., Nwaila, G.T., Tolmay, L., Frimmel, H.E., Bourdeau, J.E. (2020/2021) "Integration of Machine Learning Algorithms with Gompertz Curves and Kriging to Estimate Resources in Gold Deposits." Natural Resources Research 30(1):39–56. DOI: 10.1007/s11053-020-09750-z. Witwatersrand conglomerates; geology as predictor set.
• Erten, G.E., Yavuz, M., Deutsch, C.V. (2022) "Combination of Machine Learning and Kriging for Spatial Estimation of Geological Attributes." Natural Resources Research 31(1):191–213. DOI: 10.1007/s11053-021-10003-w. The most-cited of this group (~68 citations). Its domain handling is behind a paywall and we could not verify it — deliberately not characterised in the episode.

Hybrid geostatistics + ML / does ML actually beat kriging

• Cotrina-Teatino, M.A., Marquina Araujo, J.J., Mamani-Quispe, J.N., Arango-Retamozo, S.M., Gonzalez-Vasquez, J.A., Vega-Gonzalez, J.A. "Mineral resource estimation using spatial copulas and machine learning optimized with metaheuristics in a copper deposit." Earth Science Informatics, Sep 30, 2025. DOI: 10.1007/s12145-025-02009-2. R²=0.82, RMSE=0.12, best model KNN tuned with a genetic algorithm. Note: the paper's own headline total-resource tonnage figure doesn't reconcile with its own reported grade/volume numbers (off by roughly two orders of magnitude, likely a percent/fraction unit slip) — we deliberately did not repeat that number on air.

• Han, H., Suh, J. "Comparative Analysis of Machine Learning–Kriging Integrative Approaches for Enhanced Spatial Prediction of Mineral Exploration Data." ISPRS International Journal of Geo-Information, Apr 15, 2026. DOI: 10.3390/ijgi15040175. Tests 12 hybrid configs (6 ML backbones × Ordinary/Universal Kriging) on aluminum concentration under spatial cross-validation. Finding: hybridizing a strong backbone (Random Forest) gains essentially nothing; hybridizing weak backbones helps somewhat but not dramatically. Standalone RF ties with RF+OK/RF+UK and beats the other ten combinations.

• Germanou, M.K., Pavlides, A., Varouchakis, E.A. "Comparison of Geostatistical and Machine Learning Methods for Spatial Analysis of Natural Resources Data." Mathematical Geosciences, Dec 18, 2025. DOI: 10.1007/s11004-025-10239-9. SOM-augmented ordinary kriging vs. Gaussian process regression and other ML on zinc data, evaluating uncertainty quantification alongside point accuracy.

Domain-conditioned ML (the "don't ignore the geologist" correction)

• Maleki, M., Mery, N., Soltani-Mohammadi, S., Plaza-Carvajal, J., Varouchakis, E.A. "Integrating Geological Domains into Machine Learning for Ore Grade Prediction: A Case Study from a Porphyry Copper Deposit." Minerals, Nov 8, 2025. DOI: 10.3390/min15111175.

• Bağ, C., Frieman, B., Westman, E. "A Comparative Study of Machine Learning and Traditional Techniques for Grade Prediction and Grade-Tonnage Evaluation in a Small VMS Deposit." Minerals, Mar 7, 2026. DOI: 10.3390/min16030280. Evaluates on the full grade-tonnage curve, not just point-grade RMSE/R².

• Borah, A., Dutta, P.J., Emery, X. "Geostatistically Enhanced Learning for Supervised Classification of Wall-Rock Alteration Using Assay Grades of Trace Elements and Sulfides." Minerals, Oct 29, 2025. DOI: 10.3390/min15111128. ~8 percentage-point accuracy gain (69%→77%) from geostatistically-derived proxy features vs. raw assay features.

• Erdogan Erten, G., Boisvert, J. "Simultaneous Estimation of Categorical and Continuous Variables with DeepKriging." Natural Resources Research, online Aug 22, 2025 / print Feb 2026. DOI: 10.1007/s11053-025-10555-1. Extends the "DeepKriging" neural architecture (spatial-coordinate basis functions feeding a feed-forward net) to jointly predict domain and grade in one pass.

• Baeza, D., Maleki, M., Varouchakis, E.A. "Leveraging Pre-existing Geological Model to Generate Multiple Realizations of Geological Domain Through Machine Learning Algorithms." Natural Resources Research, Nov 21, 2025 (print Feb 2026). DOI: 10.1007/s11053-025-10572-0.

• Zhang, S., Xue, J., Liu, Y., Lin, J., Zhou, J., Zhao, J., Zhang, Y., Li, J., Zhao, F. "Bauxite Identification and Grade Prediction from Well Logs Using XGBoost: A Case Study from Shanxi Province, China." Minerals, Dec 31, 2025. DOI: 10.3390/min16010053.

Generative models moving into conditional simulation

• Han, H., Suh, J. "GSA-cGAN: A Geospatial-Aware Conditional Wasserstein Generative Adversarial Network for Mineral Resources Interpolation." Applied Sciences, Jan 8, 2026. DOI: 10.3390/app16020674. 272 samples, Taebaek Mountains, Korea.

• Xu, M., Song, S., Mukerji, T. "DiffSIM: Unconditional and conditional facies simulation based on denoising diffusion generative models." arXiv, Mar 7, 2026. arXiv:2603.07383. Reservoir-characterization framing (oil & gas), not a mineral case study — flagged as a trend to watch, not a direct result. Reports a 30× inference speedup via DDIM sampling.

• Rahimi, I. "Deep Learning Decision Support System for Open-Pit Mining Optimisation: GPU-Accelerated Planning Under Geological Uncertainty." arXiv, Nov 23, 2025. arXiv:2511.18296. VAE trained on 50,000 spatial grade samples as a fast conditional-simulation surrogate, feeding a GPU-parallel GA+LNS+SA+RL metaheuristic across 65,536 pit-optimization scenarios. Abstract claims up to a 1.2-million-fold runtime improvement over CPLEX and significantly higher expected NPV under geological uncertainty. Tested on synthetic/generic data, not a named real deposit, and not yet independently replicated — treat the magnitude of that speedup claim with real caution until someone else reproduces it.

Uncertainty quantification & reserve/material classification

• Wang, Z., Zuo, R., Kreuzer, O.P. "Uncertainty Quantification of Deep Learning Algorithms for Lithological Mapping." Mathematical Geosciences, online Oct 6, 2025 (print Feb 2026). DOI: 10.1007/s11004-025-10235-z. Integrates Bayes-by-backprop, Monte Carlo dropout, and deep ensembles into one CNN.

• Vijouyeh, A.G., Kadkhodaie, A., Siahcheshm, K., Asadi, A., Hosseinzadeh, M. "Integrated ore classification using stand-alone and hybridised machine learning algorithms." Scientific Reports, Mar 23, 2026. DOI: 10.1038/s41598-026-42248-x. Committee-machine classifier, 19 trace elements, 8 drill holes, Sari-Gunay gold-polymetallic mine, Iran.

• Leung, R., Melkumyan, A. "Effective information gathering for ore estimation, evaluation and perspectives on adaptive sampling." arXiv, May 22, 2026 (accepted IEEE ICII 2026). arXiv:2605.23172. Gaussian-process framework for valuing an additional drill hole before drilling it; adaptive sampling beats grid drilling in geologically discontinuous zones.

Spatial ML / infrastructure / adjacent

• Chauke, T. "Graph neural networks for 3D spatial continuity analysis: A comparative study with ordinary kriging in ore grade estimation." SME technical abstract, May 27, 2026. No DOI — not a full peer-reviewed paper, treat as "watch this space" rather than a settled result.

• Thiele, S.T., Kirsch, M., Frenzel, M., Tolosana-Delgado, R., Kamath, A.V., Guy, B.M., Kim, Y., Tuşa, L., Járóka, T., Gloaguen, R. "Predicting Mineralogy with Hyperspectral Data: A Benchmark Dataset and Machine Learning Framework to Enable Hyperspectral Geometallurgy." Minerals, Jun 26, 2026. DOI: 10.3390/min16070674. Open benchmark dataset + framework, infrastructure more than a single result.

• Kriuk, B. "Korzhinskii-Net: Physics-Informed Neural Network for Sub-Surface Mineral Prospectivity Modelling." arXiv, May 31 2026 (rev. Jul 1, 2026). arXiv:2606.13695. Exploration targeting, not resource estimation — mentioned as an adjacent trend (physics-informed architectures). PR-AUC 0.708 vs. 0.235 for the best classical baseline across 6 districts / 3 commodities.

Mentioned for context (published just outside the Sep 2025–Sep 2026 window)

• Erdogan Erten, G., Mokdad, K., Zacche da Silva, C., Nisenson, J., Brandao, G., Boisvert, J. "Ensemble Machine Learning Geostatistical Hybrid Models for Grade Control." Mathematical Geosciences, Jan 2025. DOI: 10.1007/s11004-024-10172-3. Direct precursor to the DeepKriging paper above.
• Li, Z., Zhan, Z., Hu, J., Yi, S., Zhang, X., Weng, Z., Zhang, Z., Ding, K. "An Adaptive Generalized Regression Neural Network Approach for Ore Grade Estimation Considering Spatial Anisotropy." Natural Resources Research, Jul 26, 2025.
• Samanta, G., Dey, T. "Advancing Grade Estimation in Mining Using Artificial Neural Networks: A Case Study of Indian Iron Ore Deposits." Journal of The Institution of Engineers (India): Series D, May 7, 2025.
• Cotrina-Teatino, M.A., Marquina-Araujo, J.J., Riquelme, Á.I. "Comparison of Machine Learning Techniques for Mineral Resource Categorization in a Copper Deposit in Peru." Natural Resources Research, May 18, 2025.

Cut from the episode

• A preprint combining concentration-number multifractal analysis with an ML grade model on a porphyry copper deposit (attributed to Yan, Zhang & Zhang, SSRN, Feb 2026 in an earlier draft) could not be independently verified to exist as described during fact-checking. Rather than cite it, the episode explicitly says on air that this claim was cut for being unverifiable.

Other sources referenced with explicit caveats on air

• A Forbes piece, "AI Can Find The Gold. The Rulebook Won't Let It Count." (reported June 2026, author Dara Abasiita) — headline used as framing; full text was paywalled and not independently verified.
• Reported (not independently confirmed) that a pending JORC Code revision will explicitly address machine learning / advanced geostatistical methods, conditional on Competent Person validation. Industry reporting only — primary committee text not published at time of research.

Glossary — ML methods and stats terms used this episode

• Kriging — geostatistics' classic spatial estimator: a weighted average of nearby samples, weighted by a variogram-based model of spatial correlation. Minimizes estimation variance, which means it smooths out local extremes.
• Variogram — a model of how similar two samples are, as a function of the distance/direction between them; the mathematical backbone kriging's weights come from.
• Conditional simulation — instead of one smoothed kriged map, generates many equally-probable realizations that honor the hard data while preserving local variability kriging averages away.
• Copula — a statistical tool that separates "how does each variable behave on its own" from "how are they tangled up together," allowing flexible dependence modeling without forcing one rigid joint-distribution shape.
• Random Forest / XGBoost / AdaBoost — tree-based ensemble ML methods; combine many decision trees to make a more robust prediction than any single tree.
• ResNet / U-Net — convolutional neural network architectures originally built for image tasks, repurposed here for spatial grade/resource prediction.
• Spatial Transformer Network — applies the "attention" mechanism behind large language models to spatial data, letting a sample point learn which other points are informative for predicting it, rather than that relationship being fixed in advance by a variogram.
• GAN (Generative Adversarial Network) — two networks trained against each other, a generator producing fake samples and a discriminator trying to catch them, until the generator produces realistic output.
• Wasserstein GAN — a GAN variant using a more stable distance metric between real and generated data distributions, reducing the training instability plain GANs are known for.
• Diffusion model — learns to reverse a gradual noising process, generating output by iteratively denoising from static; strong at preserving fine texture. DDIM is a faster sampling variant.
• VAE (Variational Autoencoder) — compresses data into a compact latent representation and decodes new samples from it; faster to sample from than diffusion, somewhat less fine-grained.
• DeepKriging — a neural architecture that converts spatial coordinates into basis-function features (so the network learns from "distance to reference points" rather than raw coordinates), letting a feed-forward network approximate and extend kriging-style prediction.
• Bayes-by-backprop — maintains a full probability distribution over a neural network's weights instead of single point-estimate weights, to produce calibrated uncertainty.
• Monte Carlo dropout — reuses the dropout regularization trick at inference time (randomly disabling neurons on repeated passes) and uses the spread of resulting outputs as an uncertainty estimate.
• Deep ensembles — train several independently-initialized networks and use their disagreement as the uncertainty signal.
• Committee machine — a classifier built from a "committee" of multiple stand-alone and hybridized models; can also serve as an implicit uncertainty signal via how much the members disagree.
• Gaussian Process (GP) regression — a probabilistic, kernel-based regression method that naturally produces a calibrated uncertainty estimate alongside each prediction (used in the Leung & Melkumyan adaptive-sampling framework).
• Physics-informed neural network — a network with real physical process equations (e.g. fluid flow, heat transport) embedded directly into its architecture, rather than learning purely from data.
• R² (R-squared) — a 0–1 score for how much of the real variation in the target a model explains; 0 = no better than guessing the average, 1 = perfect.
• RMSE (Root Mean Squared Error) — the typical size of a model's prediction error, in the original units; lower is better.
• PR-AUC — a score (0–1) for how well a classifier ranks true positives above false alarms across all thresholds; used here for prospectivity-mapping performance.
• JORC Code / 43-101 — the Australasian / North American mineral reporting codes that define Measured/Indicated/Inferred resource confidence categories and require a named, licensed Competent Person / Qualified Person to sign off on a resource or reserve statement.

Episode built with the make-a-podcast skill — script, fact-check, and audio generated by Claude, reviewed and corrected in collaboration with the listener before finalizing.