AI groups 935 iron meteorites to trace the remnants of ancient planetary cores

Iron meteorites carry chemical evidence from the early Solar System. Many are fragments of the metallic cores of planetesimals—small bodies that separated into layers before collisions broke them apart. Comparing those fragments helps reconstruct the environments in which their parent bodies formed.

At Birkbeck, University of London, researchers assembled a remarkably broad chemical archive and asked a machine-learning system to find the groupings hidden within it. Their August 2026 report describes nearly 2,500 published analyses covering 935 iron meteorites—roughly two-thirds of those then known—drawn from 61 papers produced between 1967 and 2023. Unsupervised cluster analysis reproduced familiar families and suggested that 29 meteorites previously left unclassified could be assigned to existing groups.[1][2]

A catalogue of destroyed cores

Nickel and trace elements such as gold and platinum preserve chemical differences within the metal. Classification groups similar elemental patterns as evidence of related formation histories. Establishing a particular parent body requires that chemistry to be considered alongside mineralogy and other measurements.

The traditional classification process has relied heavily on plotting combinations of elements against one another. Such plots are useful because a human scientist can see clusters and unusual cases. They also have a limitation: a series of two-dimensional views may struggle to capture relationships spread across many chemical measurements at once. Older analyses were produced by different laboratories using different techniques, so researchers must consider measurement consistency and historical gaps before trusting a neat computer-generated boundary.[1][2]

What 'unsupervised' actually means

A supervised model learns to recognise examples already labelled by people. Unsupervised clustering instead starts by asking which observations sit near one another according to a chosen definition of similarity. For meteorites, the features are elemental abundances; the model can examine many of them simultaneously. The result is a set of chemically related candidates that specialists can compare with mineralogy, existing classification and known formation mechanisms.

Clustering depends on the chosen elements, their scaling and the treatment of missing measurements. Its agreement with established families provides a useful check, while the 29 proposed assignments give specialists specific cases to investigate further.

The value of the cases that refuse to fit

Classification systems are judged partly by the ordinary cases they describe. In planetary science, the exceptions can be more valuable. If a meteorite sits well outside known chemical families and its measurements survive independent reanalysis, it might indicate an unusual formation environment or material from another parent body. Conversely, an object once called 'ungrouped' may have appeared mysterious simply because previous plots did not combine the right information.

The Birkbeck team's free release of both the method and underlying data matters here. It lets other laboratories test whether classifications are stable when a variable is omitted, compare new samples with historic ones and question uncertain assignments. Open methods are particularly useful when the evidence is irreplaceable: nobody can go back and collect additional rocks from a destroyed planet to resolve a disagreement.[1][2]

A different kind of AI discovery

The result is a revision to meteorite classification: 29 existing objects may belong to already recognised groups. Combining more measurements in one analysis can resolve cases that looked isolated in simpler plots.[1][2]

The reusable archive is part of the achievement. Other laboratories can compare new specimens against the same chemical record, test uncertain boundaries and refine the family tree as further evidence arrives.

Report an error or suggest a correction