An artist makes a quick charcoal sketch, and the face in the portrait is already recognizable. Astrophysicists face a similar challenge: the spectrum of a star is its individual portrait, revealing temperature, composition, and gravity. Until recently, obtaining such a portrait required painstaking 'oil painting'—radiative transfer codes under non-local thermodynamic equilibrium (non-LTE). One accurate synthetic spectrum took hours to compute. But in the era of new-generation surveys like WEAVE and 4MOST, millions of portraits are needed. Enter SpecFANN — a neural network trained to draw fast yet surprisingly accurate sketches of stellar spectra.

The research team took 50,147 models generated by the FASTWIND code — a kind of reference 'artist'. Each model represented a spectrum of a star with specific parameters: effective temperature, surface gravity, radius, helium abundance, and abundances of C, N, O, Si. This set became a training album. 140 independent neural networks (one per important spectral line) learned to predict the line shape from five input numbers. After about an hour of training on an NVIDIA L40S GPU, each network starts outputting 161 profile points almost instantly. The physics of the stellar core (the CNO cycle, first described by Hans Bethe) leaves characteristic signatures in the spectrum that the neural network learned to capture without direct radiative transfer calculations.

A quick sketch may blur small details, but the main features remain. For photospheric lines (those born in denser layers), the average error drops below 0.1%—almost as good as the original model. Wind lines, which arise in the rapidly expanding envelope, are trickier, but even there the error rarely exceeds 1%. Hydrogen lines, such as Hα from the Balmer series, discovered by Johann Jakob Balmer, are reproduced very well. A real surprise was that accounting for stellar rotation, a slight 'smearing' of the profile by the Doppler effect, further smooths out small prediction noise, reducing the error by half.

The speed of the emulator (360,000 times faster) turns the impossible into routine. Bayesian analysis methods, such as MCMC, which iterate over thousands of possibilities, become applicable for non-LTE models. Astronomers no longer have to settle for sparse spectral indices: they can use all the information carried by hundreds of lines. In the future, the network architecture will be optimized for different metallicities, add wind parameters, and account for binary systems. Thanks to SpecFANN, large-scale analysis of stellar evolution using survey data will take days, not centuries.

A quick sketch won't replace a detailed painting when 100% accuracy is needed. But astrophysics is entering an era where what matters are not individual masterpieces but statistical ensembles. SpecFANN is the tool that lets you quickly glance at thousands of portraits and find the most remarkable stars (future supernovae, gravitational wave sources) and then study them in full detail. A charcoal drawing gives the essential: speed. And with it comes the ability to see the forest for the trees.