From the lithium in our EV batteries to the iron used to construct new buildings, we constantly rely on minerals hidden throughout Earth. Finding these minerals is a resource-intensive process often involving satellite imagery, geochemical surveys, and insights gleaned from past extraction projects. While miners sometimes hit the natural resource jackpot, locating an economically viable material cache at the right time isn't always easy. As a result, scientists are turning to artificial intelligence to improve their odds.
Researchers at the University of Notre Dame, the University of Arizona, and the Carnegie Institution for Science have developed a model that analyzes mineralogical data to predict the locations of rare or useful materials. The model is based on data from the Mineral Evolution Database, which houses information about 5,477 unique types of minerals and their 295,583 localities. As the team writes in their paper for PNAS Nexus, the sheer volume of this data—combined with “the complexity and inherent ‘messiness’ of our planet’s intertwined geological, chemical, and biological systems”—makes it extremely difficult for humans to locate resource stores using the Mineral Evolution Database.
But AI is uniquely poised to assess this data quickly. The model’s secret sauce is association analysis, which involves searching for patterns within vast datasets. By pairing the Mineral Evolution Database with information regarding plate tectonics, the oxidation of Earth’s atmosphere, geosphere evolution, and other peripheral phenomena, the model determines where mineral stores might be found, and roughly how much of the mineral might be hidden there.
The researchers tested their model in October 2020 by requesting locations for rutherfordine, andersonite, schröckingerite, bayleyite, and zippeite. For efficiency, they set parameters in which only predictions with a 70% or higher “confidence” level would be shown. The model returned four predicted localities for rutherfordine, one of which has since been confirmed in Italy; one for andersonite, which has yet to be confirmed; one for schröckingerite, which was confirmed in Colorado; two for bayleyite, both Utah locations of which have been suspected before; and seven for zippeite, one of which has been confirmed in the Czech Republic. (As with bayleyite, four of the predicted zippeite localities had already been suspected.)
Another test run focusing exclusively on raw earth elements (REE) predicted localities for monazite, allanite, and spodumene, which are used for construction, radiation research, and batteries. Out of 15 predictions, 12 monazite localities have been confirmed; 13 out of 19 allanite localities have been confirmed; and one out of 12 spodumene localities have been confirmed.
The model is believed to have potential on other planets, too. Using what we know about the Moon’s or Mars’ geological and astrobiological histories, the model could someday help locate minerals throughout the solar system.
