As you go through this iterative process of trying to find what the telescope actually is doing during observations, you basically see the glare of the star disappearing in very bizarre ways. It gets dimmer and very deformed until at the end you get a pretty clean image and, if you’re lucky, there’s a blob there, which is a glowing planet.
This process, described by Rodrigo Ferrer-Chavez, is central to a new computational technique that allows researchers to detect exoplanets previously hidden by the intense glare of their host stars. The method, detailed in a paper published in the journal Nature, improves the James Webb Space Telescope’s (JWST) ability to image distant worlds by using machine learning to account for the observatory’s physical instabilities. Even the largest planets can be more than 10,000 times fainter than their host stars, meaning their light is often lost within the stellar glare. To manage this, researchers utilize a coronagraph, which acts as an artificial eclipse to block out the majority of the starlight.
Physics-Based Learning for Imaging Planets
The technique, called Physics-based Learning for Imaging Planets (FLIP), moves away from traditional methods that rely on comparing a target star to a reference star. Because a telescope is never in the same state twice—mirrors flex, temperatures fluctuate, and stars shift off-center—using a reference template often leaves behind residual starlight that obscures faint planets. Ferrer-Chavez, the study’s lead author, describes the process as akin to creating an artificial eclipse.
By using gradient descent, researchers optimize 106 physical parameters to simulate exactly how the telescope is behaving during a specific observation. The process begins with a simulated point spread function, which is then fine-tuned against the actual scientific data, a step necessary because variables shift between different observations. Running this on a current GPU typically requires just over 30 minutes.
Detecting Hidden Worlds
In testing the approach on JWST data, the research team successfully identified the giant planet 51 Eridani b, which had not been visible in previous analyses at specific wavelengths. The team also applied the method to observations of the star system HIP 65426 and a third star, TYC 5899, which has no known planets but exhibits high variability. Using FLIP, the team boosted the signal-to-noise ratio by a factor of 1.5 to 2, and for TYC 5899, the approach increased sensitivity by a factor of 15.
The researchers note that by reducing noise, FLIP may allow for the detection of dimmer, older, or lower-mass planets that were previously missed. The group also anticipates that by adjusting FLIP to pinpoint where it applies the most significant model modifications, they could gain valuable insights into the status of the telescope’s hardware.
Applications for Future Observatories
Financial backing for this project was provided by the Space Telescope Science Institute, the Sloan Foundation, an Amazon AI4Science Discovery Award, and the National Science Foundation. Because the software is built on a physical model of the instrument, it can be adapted for other hardware.
If image-processing software can effectively clean up stellar glare, the physical hardware might be permitted to wobble slightly more. FLIP aims to simplify the construction of NASA’s Habitable Worlds Observatory (HWO), the successor to JWST that is scheduled for a 2040s launch.
I believe our recent success was only possible because we did that exercise first.