The Top-Cut — Ep. 1 transcript

Ep. 1 — AI Meets the Block Model

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.

Traci: So I want to start with a headline I saw a few weeks ago, and I couldn't even get past the paywall to read the actual piece, so take this as a vibe rather than a source. Forbes, and the headline was — "AI Can Find The Gold. The Rulebook Won't Let It Count."

Joao: Okay, that's a great headline.

Traci: It's a great headline, and I think it's basically the whole episode in one sentence, honestly. We could almost just leave it there.

Joao: We are absolutely not doing that, we spent way too long on this. Okay, but unpack it for me, because "AI can find the gold" — found how? Found where?

Traci: So this is what we've been buried in for a couple weeks now — everything published in roughly the last twelve months, call it September to September, sitting right at the intersection of resource and reserve estimation, so the actual number that goes on a technical report, and machine learning methods being used to produce that number.

Joao: Not exploration targeting, not "here's a prospective spot to drill."

Traci: Right, we're being disciplined about that distinction the whole episode, because it's a totally different problem. This is: you've already got the drill holes, you've got the assays, and the question is what grade and tonnage do you actually report, and can a neural network do that job as well as, or better than, kriging.

Joao: Which, I have to say, before we even open a single paper — my prior walking in was pretty skeptical. Not because I don't believe ML works, I believe it works great at a lot of things. But resource estimation specifically has this seventy-year-old, extremely conservative regulatory apparatus wrapped around it. JORC, forty-three one-oh-one, Competent Persons signing their names to numbers. That doesn't feel like fertile ground for "throw a transformer at it and see what happens."

Traci: And that instinct turns out to be basically correct, but not for the reason you'd think. It's not that the methods don't work. Some of them work really well. It's that almost every serious paper this year is wrestling with exactly the tension in that headline — the modeling capability has clearly outpaced the ability to explain, calibrate, and defend what the model did.

Joao: Okay so let's actually get into — wait, before we do, I do want to flag, there's apparently a JORC Code revision in progress that's supposed to explicitly address machine learning methods.

Traci: Yeah, and I want to be really careful here because we could not get the primary committee document, it's not published yet as far as we can tell, this is industry reporting, secondhand. So hold it loosely. But directionally, everything we read this year is consistent with a regulator trying to figure out how to let this in the door without letting the rigor out.

Joao: Conditional on the Competent Person being able to actually validate and explain the model.

Traci: That's the reported framing, yeah. And that phrase, "validate and explain," is going to come back like six or seven times today, because it turns out to be the load-bearing wall for basically this entire literature.

Joao: Okay, before we get into individual papers, I want the primer, because I actually want to understand what we're comparing against. What is kriging, actually, at a level deeper than "the thing everyone already does."

Traci: Okay so — this is a genuinely fun bit of history, actually. Danie Krige, South African mining engineer, publishes a paper in nineteen fifty-one — "A statistical approach to some basic mine valuation problems on the Witwatersrand" — working the Witwatersrand gold fields, and he notices something's off with how people were estimating grade — the naive methods were biased, they overvalued the low-grade blocks and undervalued the high-grade ones.

Joao: Because of sampling error interacting with variability, or —

Traci: Right, basically. And then a French mathematician, Georges Matheron, publishes "Principles of geostatistics" in Economic Geology in nineteen sixty-three, and turns Krige's empirical observation into an actual rigorous theory — that's geostatistics as a field — and the tool that falls out of it he names kriging, after Krige. Which is why, incidentally, it's KREE-ging and not anything else — it's a man's name.

Joao: Which is, reduced to its bones, a weighted average.

Traci: A weighted average, but the weights aren't arbitrary, they come from a model of how similar two points are as a function of the distance and direction between them — that's the variogram. Two samples ten meters apart along strike in a Carlin-type system are probably really similar. Two samples ten meters apart across a fault, maybe not similar at all.

Joao: So the variogram is encoding "how far can I trust this sample to speak for the rock around it."

Traci: Exactly, and kriging uses that to build the best linear unbiased estimate — mathematically, it minimizes variance without systematically over or under-shooting. Seventy years of use, every reporting code on earth is basically built assuming this is how you do it.

Joao: Okay so what's the catch. There's obviously a catch, or we wouldn't have twenty-some papers trying to do something else.

Traci: The catch is baked right into "minimizes variance." If you're minimizing variance, you are, by construction, smoothing. Pulling extreme values toward the local average.

Joao: Which is fine if —

Traci: If you want a stable global number for a resource statement. You don't want your total reported tonnage swinging around because of a couple of nugget-effect outliers. But it's a real problem the second you care about local detail — grade control, where you're deciding literally today which truck goes to the mill and which goes to the waste dump. The actual local highs and lows are exactly the thing you need, and kriging has averaged them away by design.

Joao: Got it. And that's presumably where conditional simulation comes in, which I know is a different thing from kriging even though people mash them together sometimes.

Traci: Right, instead of one smoothed best estimate, you generate a whole family of realizations, dozens, hundreds, all honoring the same hard drill data, but each one keeping the local roughness that a single kriged map irons out. You trade one clean map for a pile of maps and the computational cost of managing all of them.

Joao: Okay. So — first real question of the day. Does machine learning, in any of its forms, actually beat kriging at its own game. Point estimates, apples to apples.

Traci: So this is where I want to start, because there's a paper that I think is honestly the most important thing we read, and it's not the flashiest.

Joao: Go.

Traci: Hosang Han and Jangwon Suh, published in ISPRS International Journal of Geo-Information back in April. And what they do is almost brute-force in a good way — they take six ML backbones and pair every single one with both Ordinary and Universal Kriging. Twelve hybrid configurations.

Joao: Hold on, "backbone" — that's the second bit of jargon in one sentence. What does that actually mean?

Traci: Fair, and it matters, because the punchline of this paper is literally the word "backbone." It just means the main workhorse model doing the predicting — the engine, before you bolt anything else onto it. Their six were Random Forest, XGBoost, AdaBoost, a ResNet, a U-Net, and a Spatial Transformer Network.

Joao: And a random forest is — I genuinely want the plain version, because I hear that term constantly and I've never had anyone define it.

Traci: Okay, plain version. Imagine you want to predict grade at a spot, so you ask a whole crowd of people and average their answers. A decision tree is one member of that crowd — it asks a series of yes-or-no questions. Is this sample deeper than four hundred metres? Is it inside the potassic zone? Is the copper above point four percent? — and follows the branches down to a guess. One tree on its own is twitchy; change the data slightly and it gives you a noticeably different answer.

Joao: So the forest is the crowd.

Traci: The forest is the crowd. You grow hundreds of trees, each on a slightly different random slice of the data and the variables, then average them. The individual quirks cancel out and the shared signal survives. That's it — that's a random forest. It's from two thousand one, Leo Breiman, and it is still the thing to beat, which is part of today's story.

Joao: And XGBoost, AdaBoost, same family?

Traci: Same family, trees again, but built in sequence instead of in parallel — each new tree concentrates on the cases the previous ones got wrong. That's called boosting. ResNet and U-Net are neural networks borrowed from image processing. So: twelve hybrid configurations.

Joao: Twelve.

Traci: Twelve, on aluminum concentration prediction, run through proper spatial cross-validation, which — quick aside, because this matters more than people give it credit for —

Joao: Wait, what's wrong with regular cross-validation? Isn't that just the standard thing you do to check a model doesn't overfit?

Traci: It's the standard thing, but on spatial data it lies to you. If you do a random k-fold split, nearby points end up on both sides of the train-test split, and nearby points are correlated with each other, so information leaks across the split and your model looks amazing on paper and then falls over the second it sees genuinely new ground.

Joao: Oh, that's — okay, that's a little alarming actually, because that seems like an easy mistake to make and a really easy way to fool yourself.

Traci: It is an easy mistake, and it is extremely common in this literature, which is exactly why I trust the papers that do spatial cross-validation properly a lot more than the ones that don't. Anyway — Han and Suh do it properly, twelve hybrids, and the finding is basically a cold shower for the "hybridize everything" instinct. Direct quote: "the effect of integration was backbone-dependent rather than uniformly beneficial." And — "increasing model complexity through hybridization does not guarantee improved accuracy."

Joao: Okay, so what won?

Traci: Plain Random Forest, not hybridized with anything — R-squared around zero point three — basically tied with RF fused to ordinary kriging and RF fused to universal kriging, and it clearly beat the other ten combinations, the XGBoost, AdaBoost, ResNet, U-Net, and Spatial Transformer hybrids.

Joao: Quick pause — R-squared, remind everyone what that actually means, since we're going to say it a lot today and I don't want to just assume everyone remembers stats class.

Traci: Sure — R-squared is just a zero-to-one score for how much of the real variation in grade your model actually explains, versus just noise it can't account for. Zero means your predictions are no better than guessing the average every time. One means it's nailing the real pattern perfectly. So that zero point three from Random Forest isn't a great score in isolation, but it's a fair, apples-to-apples score, and the point of this paper isn't "is zero point three good," it's "does adding kriging on top of it help," and the answer is no. We'll also throw around RMSE sometimes — same idea from a different angle, it's the typical size of the model's error in the original units, so a smaller RMSE means tighter, more accurate predictions.

Joao: Wait, so hold on — plain Random Forest tied with its own hybrid versions, but the fanciest architecture in the pile, the Spatial Transformer, still lost even when you fused it with kriging?

Traci: Right, and that's actually the more precise version of the finding, and I want to correct myself slightly because I said it too cleanly a second ago — it's not "Random Forest beats everything," it's "hybridizing a strong backbone like Random Forest buys you basically nothing, it was already good enough on its own, while hybridizing a weak backbone buys you a measurable but still not knockout improvement." Complexity doesn't rescue a weak backbone and doesn't meaningfully help a strong one. And I love this paper specifically because you can tell from reading it they didn't know which way it would go when they started. A lot of ML-in-geology papers, you can smell the conclusion coming from the abstract. This one actually looks like an honest experiment.

Joao: Okay, so that's a real, kind of humbling data point. Is there a counterexample? Because it feels weird to lead the whole episode with "actually, nothing works," when presumably twenty other papers found something.

Traci: Yeah, there's a great counterexample, and it's useful specifically because it's not making the opposite mistake — it's not claiming "ML wins, full stop" either. Marco Cotrina-Teatino and a team, published in Earth Science Informatics right at the start of our window, late September last year. Copper deposit. And they do something genuinely different — they use spatial copulas instead of a variogram.

Joao: Copulas — spell that out, because it's an easy word to mishear if you've never seen it written down.

Traci: C-O-P-U-L-A, copula, and yeah, it's an obscure enough term that most people outside statistics have never encountered it. Here's the plain version — normally, if you want to model how two things vary together, you pick one specific mathematical shape for that relationship and hope it fits. A copula is a trick that separates two questions: how does each variable behave on its own, and separately, how are they tangled up with each other. That separation lets you mix and match — model each variable's own quirks realistically, then bolt on almost any dependence pattern you want between them — without forcing everything into one rigid shape the way a standard variogram does.

Joao: So it's more like a flexible connector piece between two things you already understand separately, rather than one big rulebook trying to cover the whole relationship at once.

Traci: That's a good way to put it, yeah. So instead of the usual variogram machinery, they use copulas to capture how grade values at different locations depend on each other, and then they optimize the ML side — K-nearest-neighbors, in the end — using a genetic algorithm to tune it.

Joao: And?

Traci: R-squared of zero point eight two, RMSE of zero point one two — genuinely strong fit. I'm going to skip the specific tonnage number they report, actually, because when I checked their arithmetic against the grade and volume figures they also report in the same paper, it doesn't reconcile — looks like a unit slip somewhere in their own total resource calculation, off by around two orders of magnitude. Doesn't touch the model performance numbers, which check out fine, just means I don't trust their headline resource figure enough to repeat it here.

Joao: Good catch, and good instinct to just flag that instead of quietly repeating a number that doesn't add up. Okay, so the fit statistics at least sound like a win for the ML side.

Traci: It reads like one, but here's the thing I actually respect about the paper — their own conclusion isn't "copulas plus ML beats kriging, use this everywhere." It's "method selection should be deposit-type dependent." Which sounds like a hedge, but I think it's actually the correct and kind of underrated takeaway.

Joao: Walk me through why, though — like, concretely, why would the same method not just be a lightbulb that works for every deposit?

Traci: Okay, think about a Carlin-type system versus a big porphyry. Carlin mineralization is structurally and stratigraphically controlled — it can be really continuous along a favorable horizon, and then just switch off sharply the second you cross a contact, a meter away, done. A porphyry's the opposite temperament — broad, roughly concentric grade shells around the intrusive center, changing gradually over tens or hundreds of meters, but then cut through by later structures and alteration overprints that create their own local chaos on top of the broad pattern.

Joao: So completely different covariance structure.

Traci: Completely different, and a single variogram, or for that matter a single off-the-shelf ML architecture, is not going to describe both well. The tool that nails a sharp stratigraphic contact — something with a hard decision boundary, like a tree-based method — isn't necessarily the tool for the smooth broad gradient of a porphyry shell, where something like a Gaussian process might have the edge. And a porphyry actually has both patterns going at once, at different scales, which is genuinely hard for any single model, human-built or learned, to nail in one pass.

Joao: That's a really satisfying answer actually, because it means "it depends" isn't a cop-out here, it's a real geological fact about why one architecture can't rule them all.

Traci: Right, and actually — can I back up for one second, because you asked me to explain the Spatial Transformer Network a minute ago and I just breezed past it, and I don't think it's fair to drop a term like that and keep going.

Joao: Please, because I nodded like I knew what that was and I did not.

Traci: So a regular transformer — the architecture behind basically every large language model — works by letting every element in a sequence "attend to" every other element, learn which other elements are relevant to it, and weight them accordingly. A Spatial Transformer Network takes that same attention mechanism and applies it across spatial locations instead of words in a sentence — so instead of a word attending to other words in a paragraph, a sample point attends to other sample points across the deposit, and the network learns which distant points are actually informative for predicting this one, rather than that relationship being hard-coded by a variogram ahead of time.

Joao: Oh, that's actually a really direct analogy to what a variogram is trying to do, just learned instead of specified.

Traci: Exactly, that's the pitch. In principle it should be able to learn more complex, non-stationary spatial relationships than a fixed variogram model can express. In practice, in the Han and Suh paper, it didn't beat plain Random Forest. Which doesn't mean the idea's wrong, it might mean you need way more training data than a typical deposit gives you for that flexibility to actually pay off instead of just overfitting.

Joao: Which is a great real-world instance of the thing you keep saying — more flexible doesn't automatically mean better, it means more flexible, and flexibility needs data to actually cash out into anything useful.

Traci: Right, there's a third paper I want to slot in here because I think it's the most balanced comparison of the bunch — Germanou, Pavlides, and Varouchakis, Mathematical Geosciences, December last year. And the thing I like is they don't just compare point-prediction accuracy, they evaluate uncertainty quantification too, side by side, kriging with self-organizing maps for scale, versus Gaussian process regression and other ML, on zinc data.

Joao: Hold that thought on uncertainty, because I want an entire chunk of this episode on that later, it feels like the actual crux.

Traci: It is the crux, we'll get there. But — take stock of where we are after just these three papers. Han and Suh: hybridizing doesn't automatically help, plain Random Forest can win. Cotrina-Teatino: a genuinely novel copula-plus-ML approach wins on this specific copper deposit, and their own paper says "match the tool to the deposit type," don't generalize. Germanou's group: the most careful comparison, and they're treating uncertainty as a first-class output, not an afterthought.

Joao: So nobody in this set is actually claiming "ML universally beats geostatistics."

Traci: Nobody serious. That conversation, I think, is just over. What's live now is when, how much, and what you're trading away.

Joao: Can I just sit in that for a second, actually, because I think it's worth remembering how loud the opposite claim was not that long ago. I feel like every mining conference from about twenty seventeen to twenty twenty had at least one keynote about AI making geologists obsolete.

Traci: Oh, completely. There was a whole run of vendor pitches in that window that were basically "give us your assay database and our black box will out-predict your whole technical team." A lot of it, to be fair to that era, was riding the same wave that produced real breakthroughs elsewhere — image recognition, language models, deep learning was legitimately eating a lot of fields around then. It just hadn't actually been stress-tested yet on the specific, small-sample, high-stakes, heavily-regulated problem that resource estimation is.

Joao: And now, what, seven or eight years later, we've got twenty-some serious papers and basically none of them are making that claim anymore.

Traci: Right, and I think that's actually a really healthy arc for a technology to go through, honestly. Loud overclaiming phase, quiet disillusionment as people actually try to deploy it and hit the data-size and explainability walls we've been talking about, and then — this is the phase we seem to be in now — a much more boring, much more useful phase where serious people figure out specifically where it helps and build real tools around that specific boundary instead of a general-purpose promise.

Joao: The Gartner hype cycle, basically, except actually happening in front of us in real papers instead of a slide.

Traci: Almost exactly the hype cycle, yeah — and if that arc holds, I'd guess we're somewhere around the "slope of enlightenment," to use their term, for domain-conditioned ML specifically, and still pretty early on that curve for the generative modeling stuff we're about to get into.

Joao: Okay, so if that fight's settled, what's actually the interesting fight this year? Because you've been hinting there's a bigger storyline.

Traci: There is, and honestly I think it might be the whole episode. So — quick bit of history again. Go back to roughly twenty eighteen through twenty twenty-three and there's a wave of papers throwing machine learning at ore grade estimation. And the lazy version of this story — the one I half-expected to be telling you — is that they all naively dumped raw assay values into a random forest, ignored the geology completely, and this year's papers are the correction. I went back and actually read them, and that's not what happened.

Joao: Okay, so what did happen?

Traci: Take Jafrasteh, Fathianpour and Suárez, twenty eighteen, Computational Geosciences — random forests, neural nets and Gaussian processes against kriging on Sarcheshmeh, the big Iranian porphyry copper. Their baseline is coordinates-only, exactly the naive thing. But they also test adding rock type as a predictor, and they report a significant accuracy gain when they do.

Joao: So they found the geology mattered.

Traci: They found the geology mattered. And then Kaplan and Topal, twenty twenty, in Minerals — this one's the striking one. They use one model to predict lithology and alteration at unsampled locations, then feed those predicted geological labels, plus coordinates, into a neural network for grade. Coordinates alone gives them an R-squared of about zero point one one. Add the geological labels and it goes to zero point five three.

Joao: Wait, that's not a marginal improvement. That's the difference between useless and usable.

Traci: It's nearly a fivefold jump in explained variance from one change, and that change is "tell the model what rock it's standing in." Zhang, Nwaila and colleagues, Natural Resources Research, twenty twenty-one, land in the same place on Witwatersrand conglomerates — geology as inputs, kriging-inspired weighting, and it helps.

Joao: So the signal was already sitting there in the older literature.

Traci: Sitting there in plain sight, and I think that's a better and more honest story than "everyone was naive and then got corrected." What separates then from now is subtler. In those papers geology enters as a *feature* — one more column in the spreadsheet, and in Kaplan and Topal's case a column the model itself guessed at. What this year's work does is treat geology as a *constraint* — hard domain boundaries the estimate has to respect, the way domained kriging always has.

Joao: Feature versus constraint. That's the actual distinction.

Traci: That's the whole distinction, and it's worth holding onto for the rest of the episode. A feature is a hint the model may choose to ignore if the data pulls it elsewhere. A constraint is a boundary it isn't allowed to cross. One caveat on my own account here — there's a fourth paper in this group, Erten, Yavuz and Deutsch, twenty twenty-two, the most cited of the bunch at around sixty-eight citations, combining machine learning and kriging. I could not get past the paywall to see how it handles domains specifically, so I'm not going to characterise it. I'll flag it as worth your own read rather than guess at it.

Joao: And I can already feel my blood pressure rising just hearing that description.

Traci: Right, because any working geologist's gut reaction is immediate — ignore the domains and the second that model has to extrapolate somewhere geologically interesting, it's going to get it wrong in a way a person would've caught instantly. And this year's literature is basically the field correcting itself, in real time, and it's not one paper doing it, it's like four or five independent groups all arriving at the same fix without appearing to have coordinated.

Joao: Which is a pretty strong signal, honestly, when multiple groups converge on the same correction independently.

Traci: It's the strongest kind of signal in science, frankly. Okay, let's go through them, and I want to start with the most direct one. Mohammad Maleki and a team, including Varouchakis again — he's basically the connective tissue of half this literature, guy's everywhere this year —

Joao: We should just have him on.

Traci: We should. Minerals journal, November last year, porphyry copper case study. And the setup is dead simple to explain — you build the ML model, but instead of feeding it just coordinates and geochemistry, you also feed it, explicitly, which geological domain each point sits in. Potassic core, phyllic zone, propylitic halo, whatever the actual scheme is for that deposit.

Joao: So you're handing the model a piece of hard-won geological interpretation instead of hoping it reverse-engineers something like it from the assay pattern alone.

Traci: Exactly, and the question they're asking is whether that domain-informed ML narrows the gap with, or beats, kriging that's already respecting those same domain boundaries. And it does. It's not "ML versus geology" anymore, it's "does adding geology to ML help," and the answer's yes.

Joao: Okay, but — why would that be such a big deal? Isn't "give the model more information" just obviously going to help, in the way that giving anyone more information helps?

Traci: You'd think, but there's a real reason domain boundaries specifically matter so much, beyond just being one more feature — they're often exactly where the economically important transitions happen. The potassic-to-phyllic transition in a porphyry, the favorable-to-unfavorable stratigraphic contact in a Carlin system. A model with zero access to that boundary has to try to infer a sharp discontinuity purely from local sample density, and it typically smooths right across it, the same way kriging would if you didn't domain it — or honestly worse, because a geologist consciously choosing where a hard domain boundary goes is making a deliberate decision, and an unconditioned model is just guessing based on whatever pattern happened to be in the training data.

Joao: There's also — and I feel like this doesn't get talked about enough — an interpretability angle. If your prediction jumps because you crossed a labeled domain boundary, that's a one-sentence explanation to a Competent Person. "The model's respecting the contact you mapped." If it jumps for reasons buried in a few hundred thousand learned weights, good luck explaining that to anyone who has to sign their name to the number.

Traci: Right, and that loops all the way back to the "validate and explain" phrase from the top of the episode.

Joao: Okay, next one.

Traci: Cemile Bağ, Ben Frieman, and Erik Westman, also Minerals, March this year. Small VMS deposit — volcanogenic massive sulfide, gold-bearing — and this one I actually think is the sneakiest important paper in the cluster, because of what they chose to measure.

Joao: Which is?

Traci: Most papers stop at RMSE and R-squared on individual grade predictions. This one goes all the way to the grade-tonnage curve.

Joao: Why does that matter so much more?

Traci: Because the grade-tonnage curve is the thing that actually determines what gets reported as a reserve and what the project's worth. You can have a gorgeous RMSE and still completely butcher your grade-tonnage curve, if your model's systematically smoothing out tonnage at high cutoffs — which is exactly the kind of error that wouldn't necessarily show up in an aggregate accuracy number but would absolutely show up in your economics.

Joao: So what'd they find?

Traci: Ensemble ML models — lower RMSE, higher R-squared, and they captured local grade patterns better than the traditional approach. But, and this is the same refrain as the last paper, the authors are explicit that it works best paired with geological domain knowledge, not as a replacement for it. Different deposit type, different authors, same conclusion, completely independently.

Joao: That's now two for two on "actually it's about combining them, not replacing anything." Keep going, I want to see how far this pattern holds.

Traci: Third one, and this has my favorite single number in the whole episode. Abhishek Borah, Parag Jyoti Dutta, and Xavier Emery, Minerals, late October last year, on wall-rock alteration classification in a porphyry system.

Joao: What'd they do?

Traci: Instead of feeding raw assay values straight into a classifier — which is the naive move — they first run a geostatistical simulation to generate spatially coherent proxy features. Basically let the geostatistics pre-digest the spatial structure. Then the classifier works off those proxies instead of raw numbers.

Joao: And the payoff.

Traci: Roughly an eight percentage point accuracy gain, just from that one change.

Joao: Eight points is a lot for what sounds like a pretty small architectural tweak.

Traci: It is a lot, and I think it's the cleanest single illustration of the whole cluster's thesis, honestly. Geostatistics isn't the thing ML is replacing here. Geostatistics is the thing that makes the ML better in the first place.

Joao: Okay I have to ask — is there a version of this where the causality runs the other way? Like, is there a paper doing the joint thing, where instead of "geostatistics feeds ML" or "ML feeds geostatistics," they're actually estimating grade and domain together, at the same time?

Traci: There absolutely is, and it's my favorite methodology paper in the whole set. Gamze Erdogan Erten and Jeff Boisvert, Natural Resources Research, posted online right at the edge of our window, late August last year — print issue's actually not landing till February — on something called DeepKriging. And I want to be careful with credit here, because DeepKriging isn't theirs — the original is Chen, Li, Reich and Sun, on arXiv in twenty twenty, published in Statistica Sinica in twenty twenty-four. Erten and Boisvert are extending it to the mining case.

Joao: DeepKriging. Okay, sell me on the name.

Traci: So the idea is genuinely elegant — it's a neural architecture explicitly built in kriging's image. Rather than handing the network raw X-Y-Z coordinates, it first converts location into a set of overlapping "bump" shapes centered at different points in space — those are the basis functions — so the network's real input is "how close am I to each of these reference bumps," not just a bare coordinate. Feed that into an ordinary feed-forward network and it turns out to naturally approximate kriging-style spatial prediction, but with a neural net's flexibility underneath.

Joao: So it's not neural network versus kriging, it's neural network built to imitate kriging and then improve on it.

Traci: Exactly, and what Erten and Boisvert do here is extend it to jointly predict categorical variables — domain, lithology — and continuous variables — grade — in one single pass, instead of the usual two-step workflow where you domain first and then estimate grade separately within each domain.

Joao: Why is the two-step thing actually a problem in practice? Like, concretely, what goes wrong.

Traci: Errors near a domain boundary in step one just get baked straight into step two, and there's no mechanism for the grade estimation to push back on a shaky domain call. It's a one-way street. Doing it jointly, at least in principle, lets information and uncertainty flow in both directions instead of just downstream.

Joao: Which, again, maps really directly onto Measured-Indicated-Inferred classification, because that confidence rating is fundamentally about how well you know both the domain geometry and the grade inside it.

Traci: Right.

Joao: Okay, but does it actually work, though? Like, did they show numbers, or is this one of those elegant-idea papers where the elegance is doing more work than the results?

Traci: No, they show it works, on the metrics you'd expect — better performance jointly estimating domain and grade together than the standard two-step pipeline running them separately achieves. I want to be honest that I don't have one single knockout number for you the way I did with the Borah eight-percentage-point figure — this reads more like "consistently better across the comparisons they ran" than one headline statistic. Which, being a little skeptical of my own segment for a second, is also just a harder thing to fact-check from the outside without the full paper's tables in front of you.

Joao: I appreciate you flagging that, actually, because I think it's a real distinction — some of what we're covering today has one crisp, quotable number, and some of it is "the overall pattern of results favors this approach," and those probably deserve slightly different levels of confidence when we're repeating them on a podcast.

Traci: Completely agree — I'd put DeepKriging in the second bucket. Genuinely promising, methodologically clever, real results, just not a single number I'd want to bet the farm on quoting out of context.

Joao: Fair enough. Okay, and there's a companion paper worth naming — Daniel Baeza, Maleki again, Varouchakis again, Natural Resources Research, November last year, on generating multiple ML-based realizations of the geological domain model itself, conditioned on a pre-existing interpretation.

Joao: So pushing the multiple-realizations, uncertainty-aware thinking that conditional simulation normally does for grade, back one more step, into the domain model.

Traci: Exactly, because right now domain geometry is usually treated as fixed the second a geologist draws the wireframe, and this is trying to bring realization-based uncertainty into that step too.

Joao: Okay, and I know there's a more workmanlike entry in this cluster too, something about bauxite?

Traci: Yeah, Shangqing Zhang and a big author team, Minerals, end of December, using XGBoost on well-log signatures to identify ore layers and predict bauxite grade in Shanxi Province. Less conceptually ambitious than the others, but it's automating a genuinely tedious manual task — interpreting log signatures — with something auditable and reproducible, instead of relying purely on an individual geologist's read of the log.

Joao: Which, again, ties back to that "validate and explain" thread — an XGBoost pipeline trained on labeled logs is something you can actually go back and audit, in a way that's more tractable than auditing twenty years of one person's pattern recognition, even when that pattern recognition is genuinely excellent.

Traci: Not knocking the twenty years, to be clear.

Joao: No, of course not, and honestly I don't think "auditable" and "replaces the expert" are the same claim at all, even though they get conflated a lot in these conversations. You can build a fully auditable pipeline that exists specifically to check the expert's work, or to handle the boring bulk of routine logs so the expert's attention goes to the weird, ambiguous ones instead of getting spread evenly across everything.

Traci: Which is honestly probably the actual near-term deployment pattern for most of what we're talking about today, now that you say it — not "replace the geologist," and not even really "augment every decision equally," but "triage." Let the auditable, validated model handle the high-confidence, routine ninety percent, and route the ambiguous, high-stakes ten percent to the person whose judgment you're actually paying for.

Joao: Which is a much less scary sentence than "AI does resource estimation now," and also, I think, a much more accurate description of what these twenty-some papers actually support, if you read them as a set instead of one at a time.

Traci: Right, no single one of them supports "trust the model on everything." Nearly all of them support something closer to "trust the model on the routine cases, and make sure a human's actually looking hard at the cases where it disagrees with itself, or with the geology, or where the uncertainty estimate — when there is one — is wide."

Joao: Okay, I think that's genuinely a useful frame to carry through the rest of the episode, actually — triage, not replacement. Alright. Step back with me for a second, because I want to say the big thing out loud before we move on. Five papers. Different authors, mostly, though with real overlap — shoutout again to Varouchakis, guy had a year. Five different deposit types. And every one of them lands on some version of the same corrective idea.

Traci: The "AI replaces the geologist" story is dead. Nobody serious is telling it anymore. What's alive, and genuinely thriving, is "AI amplifies the geologist" — which is a much worse sentence for a press release, and a much better one for an actual technical report.

Joao: Okay, I want to shift gears completely, because there's a cluster of stuff you flagged to me earlier that I think is the most technically wild part of the whole episode. Generative models.

Traci: Yeah, this is the one where I think people's ears are actually going to perk up, because conditional simulation — remember, the multiple-realizations thing we talked about at the top — that's traditionally geostatistics' most defensible home turf. If ML's making a real play for that, that's a bigger deal than beating kriging on an RMSE leaderboard.

Joao: Okay, quick level-set for me, because I know roughly what a GAN is from, like, deepfake headlines, but not really how it'd apply here.

Traci: Sure — three different flavors show up this year, and they're worth distinguishing because they work really differently under the hood. A GAN, generative adversarial network, is two networks fighting each other — one tries to generate realistic fake samples, the other tries to catch which ones are fake, and the generator gets better and better at fooling the discriminator over training. A diffusion model does something almost the opposite in spirit — it learns to reverse a gradual noising process, starting from pure static and slowly denoising toward something coherent, which turns out to be shockingly good at preserving fine texture. And a VAE, variational autoencoder, compresses data down into a compact latent space and learns to decode new samples back out of it — fast to generate lots of samples once trained, a bit less fine-grained than diffusion.

Joao: But all three share the one property you actually want here.

Traci: Right — generate new, realistic samples that match the statistical texture of real data, instead of collapsing to a single smoothed average the way kriging does by construction. Which is exactly what you want if the thing you're trying to augment or replace is conditional simulation.

Joao: Okay, give me the papers.

Traci: First one's got maybe the best acronym of the year — GSA-cGAN. Geospatial-Aware Conditional Wasserstein GAN, for mineral resource interpolation.

Joao: "Wasserstein" is doing a lot of unexplained work in that title.

Traci: Fair — quick gloss, not the full math. Plain GANs are notoriously finicky to train, the generator and discriminator can get stuck fighting each other without actually improving. "Wasserstein" refers to a specific, more stable way of scoring how far off the generator's fakes are from the real thing, one that tends to avoid a lot of that finicky training-collapse behavior. So "Wasserstein GAN" basically just means "a GAN, but built with the training-stability fix." Same team as the kriging-hybrid paper, actually, Han and Suh, so clearly a group working this exact space from multiple angles. Applied Sciences, January this year. Trained on two hundred seventy-two samples from a mineralized zone in Korea's Taebaek Mountains, and the whole design is explicitly aimed at kriging's over-smoothing problem — preserve variance, preserve local hotspot anomalies, preserve the texture that ordinary and universal kriging both average away.

Joao: Two hundred seventy-two samples feels really small for a GAN though — aren't those notoriously data-hungry and kind of unstable to train?

Traci: Yeah, that's my one real hesitation with it. As a proof that the architecture can be adapted to spatial mineral data at all, it's a real result, benchmarked properly against kriging and other ML baselines. I'd just want to see it hold up on a much bigger dataset before getting too excited about it as a general tool.

Joao: Fair. What's next.

Traci: This one's a preprint, so bigger grain of salt, but I think it's the most technically interesting thing in the whole episode. Minghui Xu, Suihong Song, Tapan Mukerji, arXiv, March this year. DiffSIM — diffusion-based facies simulation.

Joao: Diffusion, like the image-generator kind of diffusion.

Traci: Same family, yeah, applied to honoring hard well-data constraints while filling in geologically consistent detail between wells. And there's a genuinely practical result buried in it — a thirty-times inference speedup using a faster sampling method called DDIM, compared to standard diffusion sampling.

Joao: Why would speed be the headline result rather than accuracy?

Traci: Because one of the standing objections to simulation-heavy workflows has always been cost — properly characterizing uncertainty means running a lot of realizations, and that's expensive. A thirty-x speedup actually changes the economics of how many realizations you can afford to generate.

Joao: Okay but you said "well data," that sounds like oil and gas language, not mining.

Traci: Good catch, yeah — big caveat, this is developed and validated on reservoir-characterization-style facies problems, oil and gas framing, not a mineral deposit case study. The method's directly transferable in principle to simulating geological domains in ore deposits, but nobody's published that specific validation on a mineral deposit yet — I went looking. So — flag it as "diffusion models are coming for conditional simulation" as a trend, while being honest the mining-specific proof point isn't there yet.

Joao: Okay, and you said there was a third one that actually touches my side of the world, the finance and planning side.

Traci: Oh, this is the one I actually read the full text of. Iman Rahimi, arXiv, November last year. Deep learning decision support for open-pit mining optimization, GPU-accelerated, under geological uncertainty.

Joao: Walk me through it.

Traci: So they train a VAE on fifty thousand spatial grade samples, and use it as a fast surrogate for conditional simulation — spit out orebody realizations quickly. Then they take those realizations and run a GPU-parallel hybrid metaheuristic, genetic algorithm plus large neighborhood search plus simulated annealing plus reinforcement learning, all stacked together, across more than sixty-five thousand pit-optimization scenarios.

Joao: That is a lot of acronyms for one paper.

Traci: It's a lot of moving parts, yeah, and the claim is huge runtime gains, comparable or better NPV than classical solvers like CPLEX, at a fraction of the compute time.

Joao: Okay, walk me through why this one actually matters to me, specifically, versus the grade-estimation stuff, in plain terms.

Traci: So — the textbook-correct way to understand how sensitive your project's NPV is to geological uncertainty, not price, not cost, geological uncertainty specifically, is to run your whole valuation across many different equally-probable orebody realizations, and look at the distribution of outcomes, not just a single point estimate.

Joao: Which nobody actually does at full rigor, because —

Traci: Because generating enough realizations and then re-running pit optimization and scheduling on each one is expensive enough that most shops cut corners. Low-medium-high case, maybe, instead of a proper distribution. This paper's attacking that bottleneck from both ends — the VAE makes generating realizations cheap, the GPU-parallel optimizer makes evaluating pit designs against them cheap. Stack those together and you go from a handful of scenarios to tens of thousands.

Joao: Which changes what, exactly? Like, what can you now do that you couldn't before?

Traci: You go from reporting a single NPV with a hand-wavy sensitivity table to actually reporting a real NPV distribution with defensible downside and upside cases. And it maps straight onto real option valuation too — if generating the outcome distribution is cheap, you're way better positioned to value flexibility. The option to expand, delay, high-grade in response to how the geology actually looks once you're underground, instead of committing to one static plan up front.

Joao: Can I push on this a little, though? Because "sixty-five thousand scenarios evaluated" sounds impressive, but I've sat through enough mine-planning software demos to have a reflex where a big scenario count sometimes just means "we ran a coarse search over a huge space really fast," not "we found genuinely better answers than a smaller, smarter search would've." Optimization people can make almost any number sound big.

Traci: That's a completely fair reflex, and normally I'd say the honest version of this result is "comparable quality, dramatically faster." But I actually want to correct myself here, because I underplayed the paper's own claim when I first read it. It's not "comparable to CPLEX." The abstract literally says up to a one point two million-fold runtime improvement over CPLEX, and "significantly higher" expected NPV under geological uncertainty, not just tied.

Joao: Wait — a million-fold. Say that again, because that's not a normal-sized engineering claim, that's the kind of number that should make both of us sit up.

Traci: A million-fold, yeah, up to one point two million times faster in their reported comparison, plus a real NPV improvement, not just a wash. And I think your instinct from a second ago is exactly the right one to apply to a number that size — a claim that extreme, from one arXiv preprint, on synthetic or generic data rather than a named real deposit, without independent replication yet, is precisely the kind of result that deserves real scrutiny before anyone treats it as settled, no matter how good it sounds.

Joao: Right, because "a million times faster" can also just mean "we compared against a badly-tuned or worst-case-configured version of the classical solver," which happens constantly in these comparisons and would make the gap look way more dramatic than it really is.

Traci: Exactly the thing I'd want to check in the full paper before repeating that number anywhere that matters — how CPLEX was actually configured, what convergence tolerance it was held to, whether it's an apples-to-apples runtime comparison. So — real, verified claim as far as what the paper states: massive speedup, genuinely higher NPV, not just a tie. Real open question: does that number survive someone else trying to reproduce it, ideally on a named, real deposit instead of synthetic data.

Joao: Okay, that's a much more honest way to hold that result in my head — impressive on paper, unreplicated, worth watching closely rather than repeating as settled fact. And I'll be honest, it's still the paper from this whole episode I'm most likely to go actually read start to finish, skepticism and all.

Traci: Right. But even directionally, faster uncertainty propagation doesn't just give you a prettier tornado chart, it changes what decisions are even worth formally valuing, because compute cost stops being the limiting factor.

Joao: Okay. I promised we'd come back to uncertainty quantification properly, and I think this is the moment, because it feels like it's actually the thing tying the whole episode together.

Traci: It might be the real thesis of the year, honestly. So — here's the uncomfortable fact sitting underneath basically everything we've covered so far. A resource statement isn't just a number. Measured, Indicated, Inferred — those are confidence categories. The entire JORC and forty-three one-oh-one framework exists because you're not just supposed to report a grade, you're supposed to report how sure you are, and that has to be defensible to a Competent Person putting their name and license on it.

Joao: And most ML papers report what — accuracy on a held-out test set, basically.

Traci: Right, RMSE, R-squared. Which tells you the model was accurate on average, historically, on data it happened to see. It does not, by itself, give you a calibrated confidence interval you could put in a technical report the way a kriging variance, imperfect as it is, at least attempts to.

Joao: Okay so is anyone actually solving this, or is it just the acknowledged gap everyone points at and moves past?

Traci: There's one paper that goes at it head-on with real rigor. Ziye Wang, Renguang Zuo, Oliver Kreuzer, Mathematical Geosciences, October last year, on lithological mapping. They integrate three completely different uncertainty quantification approaches into one CNN.

Joao: Which three?

Traci: Bayes-by-backprop — and since that name is doing nothing for anyone, backprop is short for back propagation, which is just how essentially every neural network learns. You show it an example, it guesses, you measure how wrong the guess was, and then you walk that error backwards through the network adjusting each internal dial a little in the direction that would've made the answer better. Do that a few million times and the dials settle somewhere useful. Bayes-by-backprop is a variant where, instead of each dial being one fixed number, it's a probability distribution — so the network carries its own uncertainty in its weights rather than pretending to be certain. Monte Carlo dropout — reuses the dropout regularization trick at inference time, runs the same input through many times with random neurons switched off, and uses the spread of outputs as an uncertainty proxy. And deep ensembles — train several independently-initialized networks, use their disagreement as the signal.

Joao: Okay, three completely different-sounding names — for anyone who just wants the one-sentence version before we go deeper, are they all basically doing the same trick underneath?

Traci: Pretty much, yeah — all three are ways of asking the same question, "if I ran this slightly differently, how much would the answer change," and using how much the answers wobble as the uncertainty number. Bayes-by-backprop builds that wobble into the network's math directly. Dropout gets it by randomly switching parts of the network off and seeing how much that shakes the answer. Ensembles get it the bluntest way possible — just train several separate networks and see how much they disagree with each other.

Joao: Okay, that framing actually makes all three click at once. Now walk me through why this is actually hard, because on the surface "just add error bars" sounds simple.

Traci: Kriging variance comes basically for free out of the same linear algebra that produces the estimate, it's built into the method. A neural net has nothing like that built in. Dropout and ensembles are both, honestly, cheap approximations to a proper Bayesian treatment that's usually computationally intractable for a network with millions of parameters. Bayes-by-backprop's closer to the real thing, but it's more expensive and, from what I can tell reading around this a bit, sometimes finicky to get to converge well.

Joao: And even once you have some notion of uncertainty out of any of the three, you still have to check whether it's calibrated, right — if the model says ninety percent confident, is it actually right ninety percent of the time?

Traci: Right, and a lot of early deep learning uncertainty work, across fields well beyond geology, produced numbers that looked reasonable on paper and turned out to be badly miscalibrated the second anyone actually checked.

Joao: Which is exactly the thing a skeptical Competent Person should be poking at before trusting any of this.

Traci: Exactly, and that loops back to the JORC thread — if that revision really does end up recognizing ML, conditional on validation, I'd bet real money calibrated uncertainty, not point accuracy, ends up being the crux of what "validation" actually means in practice.

Joao: Can I say something slightly dark about that?

Traci: Go for it.

Joao: An accurate model with miscalibrated confidence is arguably more dangerous than a less accurate model that's honest about how unsure it is. Because the miscalibrated one gets trusted right up until the moment it's catastrophically wrong, and by then you've made decisions based on it.

Traci: Yeah. That's not solved by this paper — nothing here fully solves it — but it's the first thing we read that's even asking the calibration question properly, three rigorous approaches side by side, instead of one ad hoc uncertainty number bolted onto an otherwise standard accuracy paper.

Joao: Okay, are there other papers in this uncertainty-and-classification space?

Traci: Couple more, yeah. Ali Gholami Vijouyeh and a team, Scientific Reports, March this year — a committee-machine approach, combining stand-alone and hybridized classifiers, on nineteen trace elements from eight drill holes at the Sari-Gunay gold-polymetallic mine in Iran, sorting material into ore, low-grade ore, waste.

Joao: So that's an ML-driven cutoff pipeline, basically.

Traci: Basically, and I'd argue the committee structure itself is a quiet form of uncertainty quantification — if the different classifiers disagree on a sample, that disagreement is informative, even though the paper isn't framing it in explicitly Bayesian terms the way Wang, Zuo, and Kreuzer do.

Joao: Eight drill holes feels small to me, actually, now that I've internalized the "small sample sizes are a real risk" lesson from earlier. Should I be suspicious of that number?

Traci: It's worth flagging, yeah, and I think it's a fair place to apply the same skepticism we've been applying all episode. Nineteen trace elements per sample gives you a lot of dimensions to work with even from a small number of physical holes, which helps, but eight holes is still a genuinely small footprint for training a committee of classifiers you'd want to trust broadly. I'd read this one as a real, useful proof of concept for the committee-machine idea specifically, on a specific deposit, rather than a result you'd want to generalize to "this exact pipeline will work anywhere" without re-validating it on your own site's data.

Joao: Which, honestly, is probably a reasonable default posture for almost everything we've covered today, small-sample papers or not — read the method, not the specific numbers, as the transferable part.

Traci: That's a good general rule for this entire episode, actually. The numbers are usually deposit-specific. The methodological idea — domain conditioning, joint categorical-continuous estimation, committee-based classification, information-value-driven drilling — that's the part that travels.

Joao: Okay, and there was one more you mentioned — something about drilling economics?

Traci: This one might actually be my favorite paper in the whole episode, weirdly, because it flips the question. Raymond Leung and Arman Melkumyan, arXiv, May this year, accepted to an IEEE conference. Instead of "how confident is the model," it asks "how much would another drill hole actually improve that confidence, and is it worth drilling."

Joao: Oh, that's a genuinely different question.

Traci: Totally different question. It's a Gaussian-process framework that generates learning curves estimating how much an additional sample would reduce uncertainty, without needing the ground truth at that location — which, obviously, is the whole point, because if you needed the ground truth to evaluate the hole, you wouldn't need to drill it.

Joao: And then they pair that with —

Traci: An adaptive sampling strategy, let the algorithm suggest where to drill next based on expected uncertainty reduction per dollar, and it beats plain grid drilling specifically in geologically discontinuous zones, which is exactly where grid drilling is weakest, because a fixed grid has no way of knowing it just crossed a structural break.

Joao: That's drilling economics dressed up as a machine learning paper.

Traci: I mean that as a compliment, genuinely. Infill drilling budgets are a real, constrained thing at every single operation, and a framework that tells you where the next hole buys the most uncertainty reduction, especially around known discontinuities, is actionable in a way a lot of this literature just isn't yet.

Joao: Okay, I want to do a quick lightning round on a few more things before we get to advice and takeaways, because I know there's a graph neural network thread and something about a benchmark dataset.

Traci: Yeah, lightning round's fair, because honestly the coverage here is thinner than the hype would suggest. Graph neural networks are everywhere right now in other fields, and I went in expecting a pile of GNN grade-estimation papers.

Joao: And?

Traci: We found exactly one, and it comes with an asterisk. Tiyani Chauke, presented through SME's technical abstracts in May, applying a GNN to three-dimensional spatial continuity in a structurally heterogeneous carbonate-hosted iron deposit, versus ordinary kriging.

Joao: What's the asterisk?

Traci: It's a technical abstract, not a full peer-reviewed paper with a DOI. SME member content, not a paginated journal article. As reported it claims better accuracy and better preservation of short-range variability than OK, which would track with what GNNs are generally good at — representing irregular, non-grid spatial relationships, and drillhole networks are absolutely that. But I want the full paper before putting real weight on it. Call it "watch this space."

Joao: And where are the GNNs actually showing up, if not here?

Traci: Prospectivity mapping, exploration targeting — a genuinely real and interesting trend, but a different question than the one we set out to answer today, so we're mostly leaving it out.

Joao: Okay, but why though — like, is there a structural reason GNNs would be a more natural fit for prospectivity than for grade estimation, or is it just that more people happened to point their GNN at that problem first and it snowballed?

Traci: Honestly probably some of both, but I think there's a real structural reason underneath it too. Prospectivity mapping is naturally a graph-shaped problem — you're relating discrete things to each other, known deposits, faults, geophysical anomalies, geochemical anomalies, and the relationships between those discrete entities are exactly what a graph neural network is built to reason over. Grade estimation is more of a continuous field problem — you're predicting a smoothly varying quantity across space, which is closer to what kriging and DeepKriging-style architectures are already built for. You can absolutely represent a drillhole network as a graph and there's no law against it, it's just a less obviously natural fit than the discrete-relationship structure of prospectivity, so I'd guess that's part of why the one example we found is a technical abstract and not a wave of full papers yet.

Joao: That actually makes me trust the thinness of that literature more, rather than reading it as "the field just hasn't gotten around to it yet." It sounds like there might be a real reason grade estimation resisted this particular hammer.

Traci: Right, and I think that's a healthy general instinct — when a trendy architecture is everywhere in one sub-problem and nearly absent in an adjacent one, it's worth asking whether that's a gap waiting to be filled or a sign the architecture's natural shape just doesn't match the second problem as well.

Joao: Okay, benchmark dataset.

Traci: This one I think quietly matters more than any single result in the episode, even though it's not flashy at all. Samuel Thiele and a big multi-author team, Minerals, late June this year — they release an open hyperspectral drillcore dataset plus an ML framework for predicting mineralogy, not just elemental grade, from spectral data.

Joao: Why does a dataset get its own moment next to all these methods papers?

Traci: Because open benchmarks are how fields actually accelerate. A good one lets a dozen research groups test their methods against the exact same data and compare results apples to apples, instead of every paper using a different proprietary deposit nobody else can access. I'd genuinely bet a chunk of next year's crop of papers cites this dataset.

Joao: Alright, and you mentioned a couple of wildcards.

Traci: Two quick ones. Boris Kriuk, arXiv, May, revised in July — Korzhinskii-Net, a physics-informed neural network for prospectivity, meaning actual fluid-flow and heat-transport and mineral-precipitation physics is embedded, differentiably, into the network, not learned purely from data. Tested across six districts, three commodities, and on the metric they used — basically a score for how well it ranks real deposits above false alarms — it scored zero point seven zero eight versus zero point two three five for the best classical baseline.

Joao: That's a huge gap if it holds up.

Traci: It is, and it's exploration targeting, so technically outside our core theme, but "physics-informed" is a real trend worth a sentence, because almost everything else we covered today is either pure data-driven ML or classical geostatistics, without much genuinely encoding geological process physics directly into the network.

Joao: You mentioned there was one more, something about multifractal analysis and porphyry copper?

Traci: There was a preprint I flagged in my notes on exactly that, combining concentration-number multifractal analysis with an ML grade model on a porphyry system, and I'm actually going to pull it, because when I went back to double check it, I couldn't independently verify it exists as described — no trace of it in the usual places. Given everything else today we've been able to check against a real DOI or arXiv ID, I'd rather just cut a claim I can't verify than repeat it here.

Joao: I appreciate that, genuinely, especially this late in the episode when it'd be easy to just let it slide through.

Traci: The general idea's still worth a sentence though, even without that specific paper attached to it — concentration-number multifractal analysis is a real, decades-old technique — Cheng, Agterberg and Ballantyne, Journal of Geochemical Exploration, nineteen ninety-four, the original concentration-area paper, for separating background mineralization from genuinely anomalous high-grade zones. It would be a very natural thing to combine with a modern ML grade model, attacking exactly the problem we flagged an hour ago, that kriging smooths away local high-grade zones. I just don't have a specific, verified twenty twenty-five or twenty twenty-six paper to point you to that's actually done it.

Joao: Which is a fine place to land, actually — flag the idea as promising and worth watching for, without pretending we've read a paper that proved it out.

Traci: Right, and I think that same "old tool plus new architecture" pattern is true of more of today's episode than people probably expect walking in, even in the papers we could fully verify. Kriging's seventy years old. Multifractal analysis in exploration geochemistry is decades old. Geostatistical simulation is decades old. The Borah, Dutta, and Emery result from earlier — geostatistical simulation feeding a classifier for that eight-point accuracy gain — is exactly this pattern with a paper we could actually confirm end to end. Almost none of the actual geological or statistical insight across this literature is new. What's new is compute cheap enough, and architectures flexible enough, to actually wire these mature ideas together in ways that would've been computationally impractical even ten years ago.

Joao: Which is a nice reframe of the whole episode, actually — less "AI is inventing new geology," more "AI is finally letting us afford to combine the good ideas we already had."

Traci: That's a much better one-sentence summary than anything I had planned, and now I'm annoyed you got there before takeaways.

Joao: I'll let you have it in the official takeaways, don't worry, I won't steal the moment twice.

Traci: Generous of you.

Joao: Before we get to the practical version, I want to do the thing where we answer the questions I bet people are already shouting back at us.

Traci: Go.

Joao: Okay, question one, and I think this is a real one — almost everything we've covered today implies a fairly serious dataset. Hundreds of drill holes, thousands of hyperspectral pixels for the Thiele benchmark, fifty thousand training samples for the Rahimi VAE. What if you're a junior explorer with forty holes on one target. Does any of this apply to you at all?

Traci: Honestly? Mostly no, and I think it's important to just say that plainly instead of oversell this to an audience it doesn't serve yet. Deep learning specifically is data-hungry — forty holes is nowhere near enough to train a neural network of any real complexity without massively overfitting, no matter how clever the architecture is. Random Forest and some of the classical ML in Cotrina-Teatino's copula paper can work with smaller datasets than deep learning needs, but even there you want enough samples that your spatial cross-validation folds actually mean something, and forty holes on one target is thin for that too.

Joao: So where would you actually point a data-constrained junior?

Traci: Toward the domain-conditioning insight, not the deep learning insight. Even with forty holes, if you've got a solid geological interpretation — structural model, alteration zonation — making sure that's explicitly encoded into whatever estimation method you're using, even old-fashioned domained kriging, captures most of the value the fancier papers are chasing, without needing the sample size deep learning wants.

Joao: Okay, second question, and this one's a little more adversarial toward the field, but I think it's fair, because we've both been doing this long enough to remember hype cycles that didn't pan out. How much of what we covered today actually matters in five years, versus ends up as a curiosity nobody uses?

Traci: Ooh. Okay, honestly — I think the domain-conditioned ML cluster is the safest bet to matter, precisely because it's the least flashy. It's not proposing to replace anything, it's a genuinely low-risk enhancement to workflows people already run, and low-risk enhancements to existing workflows are exactly the kind of thing that quietly become standard practice without ever getting a triumphant announcement.

Joao: And the uncertainty quantification stuff?

Traci: I'd bet on that mattering too, but on a slower timeline, because it's downstream of the calibration and regulatory questions, and those move slowly by design. That's a feature of a regulatory framework, honestly, even when it's frustrating in the moment.

Joao: And the generative modeling cluster, the GANs and diffusion stuff — that's the one I actually want your honest number on.

Traci: Fifty-fifty, genuinely. If compute keeps getting cheaper and someone does the hard work of validating a diffusion or GAN-based simulator against real, named, publicly reported deposits, rather than small proof-of-concept datasets, I think it becomes a real tool within a handful of years. If that validation work doesn't happen, or the results don't survive someone else trying to reproduce them independently, this is the thing we look back on as an interesting research direction that didn't cross into practice. I genuinely don't think we know which way that goes yet, and I'd be suspicious of anyone who tells you they do with confidence right now.

Joao: I think that's honest, and it's a good instinct to hold generally with frontier claims in this space — don't just ask "did it work in this one paper," ask "has anyone tried to break it yet, on a deposit the original authors didn't hand-pick."

Joao: Okay, before takeaways, I want the practical version. Say someone listening actually runs resource estimation somewhere, junior or major, doesn't matter — what do they actually do Monday morning? Not five years from now, this month.

Traci: I'd start really boring, actually, more boring than anything we talked about today. Before touching a neural network at all — check your validation methodology. Spatial cross-validation, not random k-fold, the Han and Suh lesson. Apply that to whatever you're already doing, kriging included, and you'll learn more honest information about how well your workflow actually generalizes than swapping in a fancier model ever would.

Joao: Okay, that's basically free.

Traci: Basically free. Second thing, also cheap — if you already have domained kriging, meaning a geologically interpreted domain model already exists, you're most of the way to trying a domain-conditioned ML approach like Maleki's group did, because the expensive part, the interpretation, already exists. Feeding that same domain info into a random forest or gradient-boosted tree as a feature is a genuinely low-cost experiment.

Joao: Third?

Traci: Third, and I think most shops should actually stop here for now unless you've got a dedicated data science team — the generative modeling stuff, GANs, diffusion, VAEs, it's exciting, but it is not mature, off-the-shelf technology. None of it. If a vendor tries to sell you a commercial tool this year claiming drop-in GAN or diffusion-based conditional simulation, ask very pointed questions about validation on deposits that look like yours, not just the one case study in the marketing deck.

Joao: And the last one?

Traci: Cheapest, highest-leverage thing on the whole list, and it's barely even a machine learning recommendation — if you've got an active drilling program, go look at the Leung and Melkumyan adaptive information-value framework. Even a rough, back-of-envelope version of "where does the next hole reduce uncertainty most," applied by an experienced geologist informed by that framing, is probably worth more to your reserve confidence this quarter than any deep learning architecture we discussed today.

Joao: I like that as the actual center of gravity for the episode. Validation methodology first, domain-conditioned simple ML second, drilling-information economics third, and treat the generative frontier as something to watch and pilot cautiously, not something to bet a reserve statement on yet.

Traci: Before we do takeaways, I do want to say the skeptical part out loud, because we've been pretty positive for an hour and a half and I don't want that to read as uncritical. What actually goes wrong with this stuff.

Joao: Please, go dark on me.

Traci: Failure mode one — garbage in, garbage out, just dressed up nicer. If your assay database has QAQC problems, mislabeled samples, standards and blanks nobody actually reviewed, a neural network learns those errors just as confidently as it learns the real signal, and gives you no way to tell which is which. Kriging has this problem too, obviously, but there's a real risk that a fancier-looking model gives people false confidence the method itself somehow cleaned up data problems it has no way of detecting.

Joao: Failure mode two?

Traci: Extrapolation past the training data. Kriging, far from your data, degrades toward the global mean in a way you intuitively learn to distrust. A neural network extrapolating past its training distribution can spit out a confident, completely wrong prediction with no obvious tell anything's gone wrong — which is exactly the scenario the uncertainty quantification work is trying to guard against, and exactly why it matters so much.

Joao: Can you make that concrete for me? Like, what does that actually look like on the ground, not in the abstract.

Traci: Sure — imagine your training data's all drawn from the upper part of a porphyry system, the parts that have been drilled out over twenty years of near-surface work. You step out and drill a deep extension, into a part of the system that behaves a bit differently, maybe you've crossed into a different alteration facies at depth that just wasn't represented in the training set at all. A kriged estimate out there is going to visibly widen its confidence, pull toward something conservative, and an experienced geologist looking at the map is going to see that widening and go "okay, we're in thin-data territory, treat this cautiously." A poorly-calibrated neural network might just confidently paint a smooth, plausible-looking continuation of the shallow pattern straight down into rock it's never actually seen an analog for, with no visual cue that anything's different about that zone at all.

Joao: And that's the scenario where somebody drills based on that extrapolated model, and the deep extension just doesn't look anything like what got promised.

Traci: That's exactly the scenario, and it's also exactly the scenario the good uncertainty quantification papers are trying to prevent — a well-calibrated model should, in principle, flag that deep extension as low-confidence the same way kriging's variance would, instead of confidently painting a picture of rock it's never actually seen anything like.

Joao: Three?

Traci: This one's more human than technical — automation complacency. If a classification pipeline like the Vijouyeh ore-waste sorter works well for months, there's a real organizational risk people start rubber-stamping its output instead of genuinely checking it, especially under production pressure. Not a critique of any specific paper, just a known pattern with automation generally, and worth building a real audit step around rather than assuming vigilance holds up indefinitely under quarter-end pressure.

Joao: And four, I'm guessing, is the regulatory one we've been circling all episode.

Traci: Four is the regulatory one. If you're the Competent Person and you can't fully explain why the model produced the number it produced, you're personally exposed if it's wrong, in a way that's different from being wrong with a well-understood industry-standard method whose failure modes are documented after seventy years of use. Not a reason to avoid the technology. A reason to be the person in the room insisting on explainability and calibrated uncertainty before anything goes in a technical report — which, encouragingly, is exactly the direction the strongest papers we read are already pushing.

Joao: Okay, two more things before takeaways, because I noticed them too and I want them on the record. Where's this research actually coming from, geographically? Because reading through everything you sent me, it's a really global list.

Traci: It really is. Han and Suh, a South Korean group, wrote three of the papers we covered, across three different sub-topics. Chinese author teams on the bauxite paper and the multifractal preprint. Iranian team on the ore classification paper. Peruvian and Latin American case studies from Cotrina-Teatino's group. Australian and Tasmanian material in the skarn and hyperspectral work.

Joao: That's a real shift from a decade ago, when geostatistics innovation specifically was a lot more concentrated — South Africa, France, a handful of traditional mining schools.

Traci: Open-source tooling and relatively accessible compute have lowered the barrier for any research group with a university affiliation and a dataset to do genuinely competitive methods work, and you can see it directly in the author list diversity across just the twenty-some papers we pulled for one episode. And underneath that — none of this happens the way it's happening right now without the collapse in GPU compute costs over the last several years. That's not a geology story, it's an infrastructure story, but it's the quiet condition enabling basically every paper in the generative modeling segment.

Joao: Can you put a number on that, even roughly? Because "compute got cheaper" is one of those phrases that's true but kind of abstract until you hear an actual before-and-after.

Traci: Roughly — training a GAN or a diffusion model of the size we're talking about in these papers would've required a university lab to either own a serious GPU cluster outright or burn through a meaningful chunk of a grant on cloud compute time, back around twenty twenty or twenty twenty-one. Today, the same class of training run is something a single well-funded PhD student can do on a couple of consumer or prosumer GPUs, or a modest cloud budget, over a few weeks instead of needing dedicated infrastructure. That's not a precise figure, I don't have an exact dollar comparison for you, but directionally it's at least an order of magnitude shift in what's accessible to an individual research group rather than a well-funded national lab.

Joao: Which tracks with everything else we've said today, honestly — a technology moving from "something only a handful of elite, well-resourced groups can even attempt" to "something a competent grad student anywhere in the world can credibly try," and the geographic spread of authors we just listed is basically the visible symptom of that shift.

Traci: Exactly, and I'd bet that trend keeps compounding for a few more years yet, which means whatever "state of the art" looks like a year from now, when we hopefully do a sequel to this episode, probably has an even wider and more surprising set of authors behind it than this year's crop did.

Joao: And the other thing — where's it actually getting published, because I noticed a pattern there too.

Traci: Yeah, a huge share of what we covered sits in one journal, Minerals, from MDPI — Maleki, Bağ and Frieman and Westman, the Zhang bauxite paper, the Borah wall-rock paper, the Thiele benchmark. Five of our core papers in one open-access, fast-turnaround journal.

Joao: Is that a good thing or a slightly worrying thing?

Traci: Bit of both, honestly. Good, in that Minerals has clearly become a real hub for this sub-field — open access, reasonably fast review, you could do a lot worse than just watching that journal's table of contents if you want to keep up going forward. Worth being clear-eyed about, in that MDPI's review rigor has a mixed reputation compared to something like Mathematical Geosciences or Economic Geology, which are slower specifically because of how much weight that slowness and depth of review carries. And then on top of that, a meaningful chunk of the most forward-looking stuff this year — the diffusion paper, the physics-informed network, the multifractal ensemble, the adaptive sampling framework — is sitting on arXiv or SSRN, not through peer review at all yet.

Joao: Which is completely normal for a fast-moving methods field, arXiv-first is just how ML publishing works now, and this corner of geology is inheriting that norm.

Traci: Right, not a reason to dismiss any of it, several of these preprints are methodologically serious. Just — track, when the peer-reviewed version eventually lands, whether the claims survive the process intact or get walked back. Good discipline generally, honestly, for any fast-moving research area.

Joao: Okay. Takeaways. I'll go first this time. Takeaway one: the "ML replaces the geostatistician" story is dead in the actual literature, even if it's alive somewhere in a conference hallway. Every serious paper this year is either explicitly hybrid, explicitly domain-conditioned, or explicitly benchmarking against kriging with real honesty about where it wins and loses — the Han and Suh result, hybridizing a strong backbone buying you next to nothing, being the sharpest example.

Traci: Takeaway two from me — the actual frontier isn't point accuracy anymore, and it probably shouldn't have ever been the main event. RMSE-versus-kriging is basically a solved genre at this point, we've seen a dozen versions of that comparison. The genuinely open problems are uncertainty a Competent Person could defend to a regulator, and generative models that preserve local texture instead of smoothing it away. That's where I'd point a research budget right now.

Joao: Takeaway three, for the finance and planning crowd specifically — watch the Rahimi VAE-plus-optimization paper and anything that follows it. If deep generative models really can serve as a fast, high-quality surrogate for conditional simulation, that doesn't just change how you estimate a resource, it changes how many pit-design and scheduling scenarios you can afford to actually run under real geological uncertainty before committing capital. That compounds into real NPV impact over a project's life.

Traci: And takeaway four, maybe the most important one if you're actually responsible for a resource statement day to day — almost nothing we covered today is download-and-deploy yet. This is frontier research. Promising, methodologically serious, not sitting in a commercial software package with a button on it. The honest advice is watch this space closely, start the conversation with your software vendors about roadmap, but the geologist with the domain knowledge is still very much the load-bearing wall here, not a fallback for when the algorithm gets confused.

Joao: Which, if you rewind all the way to the top of the episode, is basically what the domain-conditioning cluster told us directly, and what almost everything else we covered reinforced from a different angle. Geology first, machine learning as amplification.

Traci: Okay, last thing, and let's actually make each other commit instead of hedging — rank your top five papers from today. Right now, out loud, no takebacks.

Joao: Oh, that's mean, we read like twenty of these. Okay — fine. Number five, I'm going with Borah, Dutta, and Emery, the wall-rock alteration paper, purely because that eight-point accuracy gain is the single most concrete, reproducible number in the whole episode, and concrete reproducible numbers are rarer in this literature than they should be.

Traci: Good pick. Mine at five is the Leung and Melkumyan adaptive sampling paper, because it's the most immediately actionable thing we covered for anyone with a live drilling budget this quarter.

Joao: Number four, mine — Wang, Zuo, and Kreuzer, the uncertainty quantification paper. It's not flashy, but it's tackling the single biggest structural weakness in the whole field, three approaches side by side, and I think it ages the best of anything we read, because the regulatory conversation is only getting more demanding about calibration, not less.

Traci: Agreed, and I'll put that at four too then, since we're allowed to overlap. Number three for me — Han and Suh, the kriging-hybrid comparison. I keep coming back to it because intellectual honesty is undervalued in this literature, and a paper willing to report "our strongest backbone gained essentially nothing from hybridization" instead of quietly reframing the story is exactly the kind of paper that keeps the rest of the field honest.

Joao: Same at three, no notes. Number two, mine — Maleki, the domain-informed porphyry paper. Cleanest statement of what I think the actual thesis of the year is — geology first, ML as amplification — and it's a real case study, not a synthetic benchmark.

Traci: And I think we both know what's at number one.

Joao: The Rahimi paper. VAE plus GPU metaheuristic pit optimization. If those results survive independent validation, it's the one paper from this whole set that doesn't just improve an estimate, it changes the economics of how many scenarios you can afford to evaluate before committing capital — and that has the widest blast radius of anything we read this year, geology, planning, and finance all at once.

Traci: Strong list. Okay — if you're going to go chase down exactly one thing after this episode, three tops, what do we send people to?

Joao: I'd say the Han and Suh kriging-hybrid paper for the reality check, the Maleki domain-informed porphyry paper for the corrective thesis, and if you're more finance and planning side, the Rahimi paper, since that's the one with the most direct line to your actual NPV model.

Traci: And if the regulatory and Competent Person angle is what you actually care about, add Wang, Zuo, and Kreuzer as a fourth. Four papers, maybe three hours of reading, and you'll genuinely have a better handle on where this field stands than most of the hot takes floating around it, including, probably, that Forbes headline we opened with.

Joao: "AI can find the gold, the rulebook won't let it count."

Traci: Which, having now read all this, I think the more accurate version is — AI can help you find the gold with a lot more nuance than that headline gives it credit for, and the rulebook is racing, reasonably carefully, to catch up.

Joao: That's a worse headline though.

Traci: Much worse headline. That's why they don't let us write them. That's the show.

Joao: Thanks for listening. We'll see you next time.