New Software Sharpens Webb's Exoplanet Images

A new software tool has cut the noise in James Webb Space Telescope images of exoplanets. With it, researchers spotted one planet in Webb data for the first time.

Better signal, one first detection

The researchers tested the method on two stars with known exoplanets, HIP 65426 and 51 Eridani. A third star, TYC 5899, has no known planets but is known to vary a lot between observations. All three are young stars. Two of them host massive planets that still glow warm after their formation by gravitational collapse, so earlier imaging was possible.

On Webb data the signal-to-noise ratio improved by a factor of 1.5 to 2, depending on the observation. In Webb data, one of the planets showed up for the first time at a wavelength suited to detecting carbon dioxide. For TYC 5899 the sensitivity of the observations rose by a factor of 15.

What the software actually does

The tool is called FLIP, a non-acronym for Physics-based Learning for Imaging Planets. It builds on a physical model of the hardware, here the NIRCam instrument on Webb, with 106 possible parameters. The process begins with a simulated point spread function and uses gradient descent to tune it against data from a real star. On a current GPU the run takes a bit more than half an hour.

The result is then fine-tuned on the science image being studied. That last step matters, because some variables differ from one observation to the next.

Why gradient descent fits the job

Gradient descent is a mathematical technique central to training large neural networks. AI researchers have become quite successful at tuning billions of parameters this way, and that a lot of work has gone into running it fast on GPUs. FLIP puts that machinery to work on a telescope model.

Why a blur matters

Point a telescope at a single point of light and the picture shows a slightly blurry spread, because of hardware imperfections and the limits of keeping the telescope aimed. The mathematical description of that distortion is the point spread function. The function is central to this work.

There are two ways to find it. One is to image a known target and derive the function from those observations. That cuts noise, yet the telescope can drift over time, and the function also shifts from one wavelength to another. The other is to model the hardware. Modern telescopes are so complicated that such models contain too many parameters to tune with practical computing, until now.

Glare, noise and dimmer planets

Researchers want to learn what conditions are like on exoplanets, by imaging them or by capturing light that passed through their atmospheres. Stars are far brighter than the planets, so both are hard. Coronagraphs block much of a star's light, but noise from subtle changes in the telescope's configuration and pointing limits them.

Lower noise should also reveal dimmer planets, meaning older or lower-mass ones, that were missed before. The team thinks it can modify FLIP to show where it changes the model most, which could say something useful about the telescope hardware. The work appears in Nature.

Source: Ars Technica