The team trained a convolutional neural network to find the "hidden correlations" in components of high-frequency data from simulated earthquakes. The AI learned how to find patterns that could be used to infer the 'missing' low frequencies, to put it another way. The result is an algorithmic system that, in the right circumstances, can estimate low frequencies and map the underground with more accuracy than before.
This isn't ready to be used in the field. The AI is only as good as its training material, and might balk if there's a quake that falls well outside the norm. A real-world system would need to understand virtually every kind of quake and underground material. And however well simulations work, the team still has to try this beyond a lab.
Even so, the tech is promising. It could help limit humanity's effect on the planet by locating more places to store CO2. It might also help find more geothermal energy pockets and let countries avoid emissions-generating electricity altogether. For that matter, there's the simple matter of understanding more about Earth. The more detailed the underground maps, the better-equipped scientists are to explain phenomenons that might have remained mysterious.