Why AI for Mineral Targeting Failed the First Time, and What Changed
Most exploration geologists have already sat through a vendor demo for an AI prospectivity tool. The pattern is familiar, a confident number, a heatmap, and very little detail about how the result was produced.
The pattern is familiar. A confident accuracy number, a heatmap, a small number of cherry-picked validation sites, and very little detail about how the result was produced. The geologist in the room asks the obvious question, which is why a particular pixel is red, and the answer is some version of the model knows. The meeting ends politely. Nothing gets drilled on the output.
This is not a failure of AI. It is a failure of how AI was sold into a discipline where every recommendation has to be defended in front of a technical committee, a board, and eventually a drill bit.
Opacity is the central problem mining has with AI
A heatmap with no explanation is not a tool a working geoscientist can use. The discipline runs on causal reasoning, on the ability to point at a target and say which structural setting, which alteration signature, and which geochemical pathfinder argue for it. A model that produces a number without producing the inputs that drove the number is not contributing to that argument. It is asking to be trusted on its own authority.
DepoDart was built on the opposite assumption. The platform is transparent and explainable. It surfaces the connections between each prediction and the inputs that drove it, so the result is easy to justify. Our team uses that explainability during the engagement, walking through it with the client's geoscientists rather than handing over a black box. This is not a feature added late. It is the part the team deliberately set out to solve, because opacity is one of the central problems mining has with AI.
What calibrated probability actually means
A second failure mode of earlier tools is uncalibrated scoring. A heatmap pixel reads 0.92, the team treats it as a 92 percent probability of mineralisation, drills the location, and finds nothing. The score was never a probability. It was a relative ranking dressed up as a percentage.
DepoDart calibrates its predictions with Platt scaling. A point scored at 0.80 should host a deposit approximately 80 percent of the time across many such points. This sounds like a small technical detail. In practice it is the difference between a model output a geologist can budget against and a model output that has to be discounted by an unknown factor every time it is read.
Uncertainty belongs in the deliverable, not in the appendix
Every prospectivity surface DepoDart produces ships with an uncertainty map. Epistemic uncertainty is estimated through bootstrap resampling, and the variance across bootstrap models flags areas where additional data collection would most improve confidence. The most defensible drill targets are the low-variance, high-probability locations. The most informative places to add a survey line or a soil grid are the high-variance locations near the threshold.
This matters for two reasons. First, it tells the team where the model is confident and where it is guessing. Second, it turns the model into a planning tool, not just a ranking tool. Where to drill is one question. Where to collect the next piece of data is another, and the uncertainty map answers it directly.
Validation that respects geographic reality
The third failure mode is validation that does not survive contact with geology. Random holdout splits on spatial data produce optimistic accuracy numbers, because nearby pixels are almost always similar and end up on both sides of the split. The model looks better than it is.
DepoDart uses spatial cross-validation with geographic blocks, and leave-one-out validation on known occurrences where data permits. Limitations are disclosed openly. In districts with fewer than three known occurrences, validation is inherently limited, and the platform says so. The honest version of the accuracy conversation is more useful, and considerably more defensible, than the optimistic one.
Honest limitations as part of the pitch
AI prospectivity mapping does not replace geological judgment. The models identify statistically anomalous feature combinations. They do not understand geological processes. In settings with no analogues in the training data, performance degrades and uncertainty rises. Results should always be reviewed by a qualified geoscientist before they inform a drill program decision.
This is the part of the methodology that should never be footnoted. A tool that refuses to overclaim is the tool that earns its way into the workflow.
What this means for your next program
The questions worth asking any AI prospectivity vendor, DepoDart included, are not about accuracy numbers. They are about what is returned alongside the score. Is the prediction explainable down to the inputs that drove it? Is the probability calibrated, and against what? Is there an uncertainty layer, and how is it estimated? Is validation spatial or random? Are limitations disclosed in the deliverable?
Ask the questions a vendor should answer.
Every DepoDart pilot delivers calibrated probabilities, an uncertainty map, spatial cross-validation, and full data provenance, with limitations disclosed in the report.