For years, an X-ray telescope in orbit recorded how binary systems shine, where an ordinary star orbits a black hole or a neutron star. Matter from the companion falls onto the compact object, heats up to millions of degrees, and emits X-rays. The archive accumulated thousands of spectra — graphs of how much light arrives at each energy. An experienced astronomer can tell a black hole from a neutron star by such a graph, but going through the archive by hand would take far too long.

The researchers trained a neural network on these spectra. The first version received only the brightnesses at different energies — hundreds of numbers. The second received the same brightnesses plus measurement uncertainties. Both learned to guess the type of object with 90–94% accuracy. But how exactly the network makes its decision remained a mystery: to it, a spectrum is just a long string of digits.

The first is the disk temperature: how strongly the infalling matter is heated. The second is the slope of the power-law tail: it shows how quickly the brightness drops with increasing energy. The third is the ratio of the disk and tail contributions: which one dominates the glow. The fourth is how well the model describes the data. The fifth is the scatter in brightness from observation to observation. These five parameters reflect the physics of accretion — the process by which matter spirals in and falls onto the compact object.

Then they built a third neural network that saw not the spectra but only these five numbers. The accuracy stayed at the same level. This means the networks really do rely on understandable astrophysical signatures, not on random noise in the data. The method is not tied to a specific satellite: the same five parameters can be extracted from data from other X-ray telescopes, current and future. So the machine not only makes a diagnosis but also explains it in a language familiar to astronomers.