FLIP software pulls exoplanets from James Webb noise

Images

×

Photo of John Timmer arstechnica.com

Researchers have extracted known exoplanets from James Webb Space Telescope data with up to twice the signal-to-noise, using software rather than new hardware. A package the researchers call FLIP, short for Physics-based Learning for Imaging Planets, builds a physical model of the telescope and tunes it against real starlight until the noise that normally buries planets largely cancels itself out. The obstacle is arithmetic: a star is far, far brighter than any light its planets send our way, which is why most of the thousands of exoplanets catalogued to date were found indirectly, by the dip in brightness when a planet crosses its star, rather than seen.

Telescopes carry coronagraphs, masks that block a star's light so nearby planets become visible. Their effectiveness runs into a wall of noise from subtle variations in the telescope's configuration and pointing between observations. A point of light never images as a point; imperfections in the hardware and limits on pointing smear it into a blur, described mathematically as the point spread function. Two standard fixes exist. One observes a known target and derives the function from real data — effective, but the hardware's behaviour changes over time and the function varies with the wavelength being imaged. The other models the hardware from engineering schematics, but modern telescopes are so complicated that the models contain too many parameters to tune.

The way through came from artificial intelligence research, which has spent years tuning neural networks with billions of parameters by gradient descent, heavily optimised for graphics cards. FLIP applies the same machinery to a physical model of Webb's NIRCam instrument, starting from a simulated point spread function, optimising against data from an actual star, then fine-tuning on the scientific image itself — a necessary final step, since some variables differ between observations. The optimisation runs on a current GPU in a bit more than half an hour.

The team tested the method on three young stars: HIP 65426 and 51 Eridani, whose massive, still-warm planets had been imaged before, and TYC 5899, which has no known planets but varies sharply between observations. Signal-to-noise improved by a factor of 1.5 to 2 depending on the data. At a wavelength that can be used to detect carbon dioxide, one of the planets turned up in data the telescope had already taken.