A star shines, and its beam, passed through a prism, scatters into a colorful stripe — a spectrum. Dark and bright lines on this stripe tell us about the star’s temperature, composition, and motion. In the past, to create such a portrait on a computer, one had to spend hours solving complex equations of light transport. It was like crafting an oil painting — slow but precise. But upcoming sky surveys are expected to produce data on millions of stars, and the old method can’t keep up.

A team of astronomers applied machine learning: they took existing numerical models — a vast album of samples. Each model is defined by five key numbers: temperature, surface gravity, the fraction of helium, and two important elements (carbon and nitrogen). The neural network learned from these numbers to predict the exact shape of narrow lines in the spectrum. An hour of training on a graphics card — and it remembers all 140 key features. Now, one star portrait is generated in a fraction of a millisecond.

The speed increased by a factor of 360 000. This made it possible to use methods that were previously out of reach — for example, rapidly sifting through thousands of hypothetical stars to find the best parameters. The Doppler effect from stellar rotation even helps: it slightly smears the lines and thereby smooths out noise, cutting the error in half. The hydrogen Hα line from the Balmer series, discovered by Johann Jakob Balmer, and traces of nuclear fusion (the CNO cycle, described by Hans Bethe) are recreated with an accuracy better than 0.1%. The neural network doesn’t even need to delve into the physics of the reactions — it simply memorizes the pattern.

A quick charcoal sketch won’t replace a finished portrait when maximum detail is required. But for large-scale studies of stellar evolution and metallicity (the content of heavy elements), it’s more important to see thousands of portraits at once. SpecFANN makes that possible — a forest of stars emerges in just days.