Variational autoencoders are applied to automatically identify galaxy populations from open high-redshift spectra obtained by the James Webb Space Telescope (JWST). Unlike traditional methods, classification is carried out without any a priori information about object types — the algorithm independently reveals the hidden structure of the data. As a result, several astrophysical classes are clearly separated, including both known and unusual, previously undescribed types of galaxies. The experiment demonstrates that unsupervised machine learning methods can automatically discover new scientific patterns in large spectroscopic surveys, paving the way to large-scale discoveries without human intervention.
The James Webb Space Telescope gathered thousands of spectra from distant galaxies—colorful fingerprints of their light. To avoid drowning in this diversity, astronomers tasked a program with sorting through the finds. But it acted like a naturalist on an unfamiliar shore: not knowing beforehand what species might appear, it arranged objects by their outward similarities. Instead of shells, it used spectral patterns. The algorithm itself identified families: the usual spirals and ellipticals, and among them, rare “blue pearls.”
Such an “automated biologist” will speed up archival sorting by hundreds of times and fish out anomalies worth a closer look. And so far, the first catch from Webb proves: even in cosmic routine, wonders are hiding.
🎯 The James Webb is so sensitive in the infrared that it could detect the heat of a bumblebee on the surface of the Moon.