There's a lot hidden underground: water, ore, sand, clay. To figure out what's there, you don't have to dig. Just run a weak alternating current through the ground and listen to how the rock 'responds' at different frequencies. This technique is called spectral induced polarization. The response reveals the composition: clay and sand, for instance, react differently.

But the response isn't just a single number: it's a whole spectrum, like a chord made of many notes. Previously, each 'note' was analyzed separately, and the overall melody got lost. Now scientists have applied a neural network that listens to the entire chord at once and immediately tells you what the ground is made of. The network was trained on 140 samples: mixtures of grains of different sizes, pieces of ore, and even cement.

The result was more accurate than previous methods: the signal reconstruction error dropped severalfold. Moreover, from the network's output, you can predict the grain size in the rock and the amount of conductive particles. Agreement with real measurements is up to 98%. And the network even learned to sort samples into categories on its own: similar soils end up close together in its internal space. This helps geologists compare new data with known ones more quickly.