Birkbeck researchers develop new AI method to unlock secrets hidden in the cores of ancient planets

Scientists at Birkbeck, University of London have developed a new machine-learning method for classifying iron meteorites - rare fragments of the metallic cores of planets that broke apart billions of years ago.

Iron meteorites are believed to be surviving fragments of the cores of "planetesimals": small, planet-like bodies that formed in the early solar system before being destroyed in collisions. As it is not possible to directly sample the core of an intact planet, these meteorites provide the only available window onto the composition of a planetary core.

Each ancient core has a distinct chemical fingerprint, shaped by the mix of iron and other metals it contains, including nickel, gold and platinum. Matching meteorites to their parent bodies has traditionally relied on manually plotting data on two-dimensional graphs, a slow process that can struggle to account for subtle chemical trends and natural variation within a single meteorite.

Louis-Alexandre Lobanov, a PhD researcher based at Birkbeck's School of Natural Sciences and the Natural History Museum, working under the supervision of Professor Hilary Downes at Birkbeck, has developed a new computational approach that classifies iron meteorites using significantly more chemical information at once. The pair built a database of nearly 2,500 published chemical analyses covering 935 iron meteorites - roughly two-thirds of all those known to science - drawn from 61 research papers published between 1967 and 2023.

Applying a machine-learning technique called cluster analysis to this dataset, the researchers were able to sort meteorites into groups automatically and consistently, based on patterns across multiple elements rather than just two at a time. The method confirmed the reliability of existing meteorite groupings and identified 29 previously unclassified meteorites that are likely to belong to known groups.

The tool and its underlying dataset have been made freely available, allowing laboratories worldwide to verify existing classifications, classify newly discovered meteorites, and potentially identify new groups, establishing a new standard for how iron meteorites are classified.

Louis-Alexandre Lobanov said:

"Iron meteorites are the only pieces of planetary cores we can study directly, but classifying them has always been a slow, manual process. Cluster analysis allows us to draw on far more of the chemical data at once, making classification faster, more reproducible and more reliable. It also means we can revisit meteorites whose classification has previously proved difficult to resolve."

Professor Downes added: "The issue of meteorite classification being a slow and clunky process is something I have been thinking about for many years. It’s a pleasure to see one of our own researchers and students be the one to solve it.”

The research, ‘An unsupervised machine learning approach to iron meteorite classification’, is published open access in Meteoritics & Planetary Science.