PNRPU Researchers Use Neural Networks to Cut Gold and Copper Ore Search Time in Half

Perm National Research Polytechnic University (PNRPU) says its researchers have taught neural networks to find ore deposits about twice as fast. The approach is designed for finding gold, copper, aluminum and other strategic metals. For the country this is an expensive line item: according to a CNews report, searching for new deposits costs hundreds of billions. Half the search time means significant savings, if the figure holds up under scrutiny.

Where exploration time goes

Finding a deposit is a chain of steps. Teams first collect geophysical, geochemical and geological data, then interpret it and pick out promising areas. Only after that do they drill. Interpretation usually relies on specialists' experience and takes weeks or months. Drilling is the most expensive step, and every borehole in barren rock eats into the budget.

Neural networks act as a filter here. They look for patterns in heterogeneous data that a human might not notice right away, or might notice too late, and they rank sites by the probability of a find. Global mining companies have been taking this route for several years. Russian work in this area is still less common, which is why the news from Perm is interesting.

What is known and what is not

The available material contains few technical details. Here is what can be stated on its basis:

• Developer: PNRPU, Perm
• Claimed effect: search is about 2x faster
• Target metals: gold, copper, aluminum and other strategic metals
• Scale of the problem: the cost of searching for new deposits runs to hundreds of billions
• Not disclosed: model architecture, the volume and source of training data, accuracy and false-positive rate, field-test results

The last item matters more than the others. Without it, "twice as fast" remains a promise, not a metric.

How to read the "twice as fast" figure

Speedup can be measured in different ways. If it refers to the time needed to process and interpret data, the result is quite realistic: a model computes faster than an analyst, which is expected. If it refers to the time from the start of work to a confirmed find, the claim is far more ambitious. That timeline depends on drilling, licensing and logistics, and a neural network has almost no effect on them.

There is a second caveat: what the model was compared against. A twofold gain over manual interpretation and the same gain over classical statistical methods are very different results. Ore problems also bring one more difficulty. Few deposits are known, training sets are small, and models tend to overfit to a specific region and transfer poorly to different geology.

What to check before adoption

For mining companies and investors, the practical sequence looks like this. First, ask the developers for a description of the method and the data the model was trained on. Then test it on a held-out site where the outcome is already known, and compare the prediction with the facts. It is important to look not only at the share of deposits found but also at the number of empty sites the model flagged as promising, because those determine the cost of an error in drilling.

If the test passes, the speedup will deliver real savings in the early stages of exploration. If it does not, the university will be left with an interesting prototype, and the market with one more number from a press release. A published paper with metrics, or a pilot with a subsoil user, will quickly show which of the two is the case.