Half an exabyte in ten years: why astronomy no longer happens from machine learning

Classification of galaxies, sorting of SETI signals, reconstruction of black hole images: astronomer Chris Impey details the uses that have become routine in front of the data flood.

The Vera Rubin Observatory in Chile will produce about half an exabyte of data over ten years. No human team can inspect such a volume frame by frame, and it is precisely this change in scale that has put machine learning at the heart of astronomical work, as Chris Impey, professor of astronomy at the University of Arizona, reminds us.

The uses are now commonplace. Neural networks classify galaxies by morphology with an announced accuracy of around 98%, and identify planetary transit signatures with about 96% success rate — a tool that has contributed to the current catalog of some 5,300 confirmed exoplanets. In the search for extraterrestrial signals, the same methods are mainly used to rule out false positives of terrestrial origin.

AI is also involved downstream of the observation: the processing of data from the black hole of M87 has made it possible to reconstruct a much clearer image than the one initially published. In terms of bibliography, the astroBERT language model indexes some fifteen million scientific articles to help researchers find their way through a literature that has become impractical to read exhaustively.