Four billion years ago, the surface of the Earth was a battleground of chaos lava flows, meteorite impact, and tectonic writhing constantly destroyed new skin of the planet. But in all this geologic brutality, a grain-sized crystal survived. Zircon, formed in molten rock and immune to heat, pressure, and chemical attack, was a minuscule time capsule. And now, those crystals, others more than 4.3 billion years old from Australia’s Jack Hills in the west, are revealing scientists an unprecedented glimpse of the vanished crust of the Hadean eon thanks to artificial intelligence.
Zhejiang University researchers, led by Professors Jia Liu and Qunke Xia and including PhD researchers Denggang Lu, Zhikang Luan, Jingjun Zhou, and Tianting Lei, created the largest geochemical dataset of its type: more than 14,000 zircon samples matched with 823 host rocks worldwide. With supervised machine learning, they taught the algorithms to learn to recognize patterns between the trace element geochemistry of zircons and the rocks they formed in. “We don’t know what the actual rocks of the Hadean crust looked like, because we don’t have any,” Lu said. “But zircons give us a window into that unseen world.”
The reconstructions defied assumptions. Rather than an andesite-like crust common to subduction zones, the initial crust is found to have been made up largely of felsic rich in quartz and feldspar and dominated by tonalite, trondhjemite, granodiorite (TTG), and potassic granites. Chemical modeling of the parent magmas suggests 58–78 wt% SiO₂, 0.1 to 1.2 potassium-to-sodium ratios, and 1 to 103 strontium-to-yttrium ratios. These values refer not to oceanic subduction far out at sea but to impact-thickened crust, since collision of the continents provided the heat and pressure required to produce felsic magmas.
Petrogenetic models account for these TTGs and granites by partial melting of mafic proto-crust dark iron- and magnesium-rich rock at several depths. Potassium-rich sources in some instances produced granites, whereas sodic compositions produced TTGs in others. This is consistent with field and experimental observation that TTGs can be produced from hydrated mafic rocks at garnet-stable, plagioclase-free conditions without modern-style subduction. However, high-pressure TTGs, such as the Aktash gneiss complex, need to melt under 60 km depth approximations, so even in the Hadean there would have been local convergent tectonics.
AI reconstructions are supplemented by other machine learning research characterizing zircon types. Zircons of Jack Hills have been discovered through studies to contain a maximum of 35% S-type, with sediment sources attributed to continental collision and crustal thickening. This high proportion suggests exposed continents, weathering, and recycling of sediments into the magmas were already established at 4.24 Ga, as is in accord with early subduction-controlled tectonic cycles. This is supported by isotopic data e.g., δ¹⁸O value variations that the Hadean Earth potentially possessed Earth surface conditions and tectonic styles similar to the present day.
The import of these results is enhanced by the paucity of direct rock records prior to 4.03 Ga, an interval commonly referred to as Earth’s “missing chapter.” Through the predictions of bulk and trace element magma compositions from zircon chemistry, the machine learning algorithms successfully reach back in the rock record by almost 400 million years. Predictions correlate with laboratory analyses of corresponding actual rock samples and show the strength of the technique.
From an engineer’s point of view, the research demonstrates the way that AI can integrate vast, high-dimensional geochemical databases to break problems previously deemed intractable. Algorithms tuned to nuanced, non-linear relationships between mineral chemistry and rock composition enabled reconstructions that could not be duplicated using standard statistical techniques. This promises to allow the application of identical methods to other early periods or even to planetary targets such as Mars and Venus, where direct sampling is restricted.
The more general geological consequences are significant. If the initial plate convergence existed, then models of continental crust formation would need to be recalculated and hypotheses concerning the timing of the emergence of life would be impacted. Volcanism associated with such tectonics may have released gases that prepared the early atmosphere, and weathering of young continents would have supplied sediments to ocean basins and started long-term geochemical processes.
Lastly, the marriage of zircon’s durability and machine learning’s ability to screen has turned these small crystals into historians of Earth’s birth. What was previously an unthinkable page in planet history is being unraveled line by chemical line into a vocabulary that algorithms can speak and analyze.
