How to Evaluate Mineral Exploration Technology: A Due Diligence Guide for Mining Investors

A junior miner pitched me last year claiming their AI could find copper deposits with 92% accuracy. I asked for the validation dataset. They sent a PDF. No raw data, no holdout samples, no comparison against drill results. Just a pretty chart.

I passed.

That meeting stuck with me because it's basically every mineral exploration tech pitch I've heard since 2022. The packaging has gotten slicker. The math? Still hand-waved. And honestly, if you're an LP writing checks into exploration funds — or a mining exec evaluating which vendor to pilot — the gap between marketing claims and defensible science is wider than most people admit.

So here's how I'd actually run due diligence on an exploration tech vendor, based on conversations with geologists, fund managers, and a few founders who've built the real thing.

Start with the physics, not the pitch deck

Every exploration technology rests on some physical signal. Spectral reflectance. Magnetic anomalies. Gravity gradients. Induced polarization. Seismic. Geochemical signatures in soil or vegetation.

If a vendor can't explain — in plain language — what physical signal their system measures and why that signal correlates with mineralization, walk away. I'm serious. I've sat through demos where the founder couldn't answer "what does your model actually see?" without retreating to "the AI handles that."

The AI doesn't handle that. The physics handles that. The AI handles pattern recognition on top of the physics.

A useful question I borrowed from a geophysicist friend at Rio Tinto: "What's the depth penetration of your primary signal, and how does that compare to the typical depth of the deposits you claim to find?" If they're using Sentinel-2 multispectral data (which only reads surface mineralogy down to a few millimeters) and claiming to find deposits at 300 meters depth, you've got a problem. They're inferring, not detecting. That's fine — but they should say so.

Companies like GeoMine AI are upfront about this distinction. Their spectral analysis identifies surface and near-surface alteration halos — the geochemical fingerprints left by hydrothermal systems — and uses those as proxies for what might sit below. That's a defensible methodology. "Our AI finds gold" is not.

The validation question nobody wants to answer

Here's where most due diligence falls apart. Investors ask "how accurate is your model?" Vendors answer with a number. Everyone moves on.

Wrong question.

The right question is: accurate against what, measured how, on which dataset, compared to what baseline?

A model that's 87% accurate at distinguishing mineralized from non-mineralized pixels in a region where 5% of pixels are mineralized is performing worse than a coin flip if you account for class imbalance. I've seen vendors quote precision without recall, recall without precision, AUC scores on training data, and — my favorite — accuracy scores calculated on the same drill holes used to train the model.

What you actually want to see:

A holdout test region the model never saw during training. Ideally a different geological province. Drill-confirmed ground truth (not just geochemistry, not just remote sensing — actual core). Performance metrics that include false positive rates, because in exploration, a false positive costs you a $2M drill program. And a comparison against the baseline a competent exploration geologist would produce using conventional methods.

If the vendor can't produce this, they either haven't done it or they did and the numbers weren't flattering. Both are red flags.

What to actually ask in the data room

I'll skip the obvious stuff (cap table, IP, team) and go straight to what I think matters for exploration tech specifically.

Ask for the training data sources. Public datasets like USGS, ASTER, Sentinel, and Landsat are great — but if that's all they've got, the model's competitive moat is thin. Anyone can retrain on the same data. The vendors with durable advantages have proprietary ground-truth datasets, often from partnerships with mining companies that gave them historical drill results in exchange for equity or licenses.

Ask how they handle the cold-start problem in new geological provinces. A model trained on porphyry copper systems in Chile may perform terribly on porphyry systems in Mongolia, even though the deposit type is similar. Vegetation, weathering, atmospheric conditions, and host rock alteration all shift the spectral signatures. Good vendors have transfer learning approaches or province-specific calibration. Bad vendors pretend the model generalizes everywhere.

Ask about field validation programs. Has any junior or major actually drilled a target the model identified? What was the hit rate? What's the cost per discovery compared to conventional methods? If the vendor has been around three years and can't point to a single drill program informed by their tech, that tells you something.

And ask — this one matters — who owns the targets the model generates. Some vendors claim equity in any discovery. Some license the software. Some operate as data providers. The commercial structure tells you whether they actually believe their own tech.

The honest answer about what this tech can do

I used to think exploration AI would compress the discovery timeline by an order of magnitude. After two years of paying close attention, I think the honest answer is more modest: good exploration technology can probably cut early-stage target generation time by 40-60% and reduce drilling costs per discovery by maybe 25-35% in well-suited terrains. That's still enormously valuable. A junior that finds a deposit two years faster on half the exploration budget is a much better investment.

But it doesn't replace geologists. It doesn't eliminate drilling. And it doesn't work equally well everywhere — heavily vegetated terrains, deeply weathered regoliths, and post-mineral cover all degrade performance, sometimes severely.

The vendors who acknowledge these limits are the ones worth backing. The ones who don't, in my experience, are usually selling a story rather than a tool.

What would you ask a vendor that I haven't covered here?