How it works

SPHEREx is a NASA space telescope that maps the whole sky in infrared light: the colours just past red, which our eyes cannot see. Cool, faint objects such as brown dwarfs give off most of their light there.

Its camera has six detectors. Each sits behind a filter whose colour changes from one edge to the other, so a single picture holds 17 strips of slightly different colour. As the telescope steps across the sky, a star passes through every strip, and together the six detectors measure 102 colours between 0.75 and 5 micrometres (millionths of a metre).

Each pixel covers 6.15 arcseconds^? of sky, about a three-hundredth of the width of the full Moon.

SPHEREx comes back to the same patch of sky about every six months. Each visit lasts from a day to a few weeks; astronomers call it an epoch^?. During a visit the telescope takes dozens of pictures of the spot, most of them through a different strip of colour.

For every visit and detector, those pictures are stacked into one cleaner image. The faint glow of dust between the planets is subtracted, damaged pixels are left out, and the rest are averaged, with the least noisy pictures counting most. One stacked image per visit is what blinks in the Lab.

These are Barnard’s Star’s real visits.

The oldest way to find things that move is to flip between two pictures of the same stars taken at different times. Distant stars hold still; anything nearby that moved jumps. Clyde Tombaugh found Pluto like this in 1930.

Subtracting one picture from the other makes it plainer still. In a difference image^?, everything that did not change cancels to grey, and a star that moved leaves a dark spot where it was and a bright spot where it is now.

Barnard’s Star, 6 light-years away, moved 10.36 arcseconds north between these two visits, as Gaia predicts to within 0.4%.

To measure how far a star moved, the exact shape a star makes on the detector is fitted to every single exposure. That pins its position down to a small fraction of a pixel. A straight line through the positions, after removing the small yearly wobble caused by Earth’s orbit, gives its proper motion^?: how fast it drifts across the sky.

The same fit measures how much light the star sent, so its brightness can be compared between visits at the same colour. RW Aurigae, a young star, faded by 0.86 magnitudes^? in a year at 3.8 to 4.4 micrometres, to less than half its brightness.

Then the answers are checked against the best catalogues. The challenge named five test stars, and the pass marks were written down before the measurements ran. Not everything passed: a failure is in the table, and the other limits are in step 7.

Because each exposure sees a star through a different strip of colour, the hundreds of exposures from a year of visits add up to a spectrum: how bright the star is at each colour.

A spectrum tells you what something is made of. Epsilon Indi B is a pair of brown dwarfs, objects too small to burn hydrogen like a star, about 12 light-years away. Water and methane in their cold atmospheres swallow whole bands of light: dips 80% and 76% deep, measured at 27 and 66 times their uncertainty.

Looking at every patch of sky by eye would take forever, so two small neural networks help.

The mover finder looks for spots that changed in difference images. It learned from fake movers and fake brightness changes planted in real SPHEREx images, where the right answer is known, and it is tested on patches of sky it never saw.

Its scores are calibrated, so a score of 80% means that about 80 in 100 such spots are real. A statistical test then sets the bar for the whole list: only candidates that keep the false-discovery rate^? at 5% are reported, so on average no more than 1 in 20 of them should be a false alarm.

The spectral classifier reads a spectrum and says what kind of object it looks most like, from hot stars to brown dwarfs, galaxies and quasars. It learned from real SPHEREx spectra of objects that catalogues had already classified.

Everything here was measured, and some of it did not work. These are the limits we know about.

Everything is public: the images, the code and the reports behind every number on this page.