Why AI for Mineral Targeting Failed the First Time, and What Changed

The pattern is familiar. A confident accuracy number, a heatmap, a few hand-picked validation sites, and very little detail about how the result was made. The geologist in the room asks the obvious question, which is why one pixel is red, and the answer is some version of 'the model knows'. The meeting ends politely. Nothing gets drilled.

This is not a failure of AI. It is a failure in how AI was sold to a field where every recommendation has to be defended in front of a technical committee, a board, and in the end 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 field runs on cause-and-effect reasoning, on being able to point at a target and say which structure, which altered zone, and which chemical clue support it. A model that gives a number but not the inputs behind it adds nothing to that case. It is simply asking to be trusted on its own word.

DepoDart was built on the opposite idea. The platform is transparent and easy to explain. It shows the links between each prediction and the inputs behind it, so the result is easy to justify. Our team uses this during the project, walking through it with the client's geoscientists instead of handing over a black box. This was not added at the end. It is the problem the team set out to solve from the start, because a lack of transparency is one of mining's main problems with AI.

What calibrated probability actually means

A second way earlier tools failed is uncalibrated scoring. A pixel reads 0.92, the team treats it as a 92 percent chance of mineralisation, drills there, and finds nothing. The score was never a real probability. It was just a ranking dressed up as a percentage.

DepoDart calibrates its predictions using a method called Platt scaling. A point scored at 0.80 should hold a deposit about 80 percent of the time, across many such points. This sounds like a small technical detail. In practice it is the difference between a result a geologist can plan a budget around and a result that has to be marked down by some unknown amount every time it is read.

Uncertainty belongs in the deliverable, not in the appendix

Every prospectivity map from DepoDart comes with an uncertainty map. We estimate this uncertainty by training the model many times on different samples of the data and measuring how much the results vary. Where they vary a lot, more data would help most. The strongest drill targets are the points with high probability and low variation. The best places to add a survey line or a soil grid are the points with high variation near the cut-off.

This matters for two reasons. First, it shows 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 gather the next piece of data is another, and the uncertainty map answers it directly.

Validation that respects geographic reality

The third way tools fail is testing that does not hold up against real geology. With map data, splitting points at random gives accuracy numbers that look too good. Nearby pixels are almost always alike, and they end up on both sides of the split. So the model looks better than it is.

DepoDart tests the model on separate map blocks, and, where the data allows, by holding out one known deposit at a time. We state the limits openly. In a district with fewer than three known deposits, testing can only go so far, and the platform says so. The honest version of the accuracy story is more useful, and far easier to defend, than the rosy one.

Honest limitations as part of the pitch

AI prospectivity mapping does not replace geological judgment. The models find unusual mixes of features in the data. They do not understand how geology works. In settings unlike anything in the training data, they perform worse and uncertainty goes up. A qualified geoscientist should always review the results before they guide a drill program.

This is the part of the method that should never be buried in a footnote. A tool that refuses to overclaim is the one that earns a place in 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 comes with the score. Can the prediction be explained down to the inputs behind it? Is the probability calibrated, and against what? Is there an uncertainty layer, and how is it worked out? Is the testing spatial or random? Are the limits stated in the deliverable itself?