Researchers have developed a machine-learning method called FLIP to improve the James Webb Space Telescope’s ability to detect and analyze exoplanets. By combining a physical model of Webb’s NIRCam instrument with gradient-descent optimization on modern GPUs, the method can be trained in just over 30 minutes and has increased the signal-to-noise ratio of observations by up to a factor of two—and, in one especially variable target, by a factor of 15.
Gradient descent is a mathematical optimization technique that has been used successfully to tune the billions of parameters in neural networks. Because it is central to training large-scale AI systems, researchers have invested heavily in making gradient descent run efficiently on GPUs.
The team’s process, called FLIP—a non-acronym for “Physics-based Learning for Imaging Planets”—uses a hardware-based physical model, in this case one representing the James Webb Space Telescope’s NIRCam instrument.6 Researchers select possible parameters and use software to optimize them. The process begins with a simulated point-spread function, which describes how the telescope images a point source, and then applies gradient descent to optimize the model using data from real stars. On modern GPUs, the procedure takes just over 30 minutes.
Researchers then fine-tune the results using the scientific images they are analyzing. This final step is important because some variables change from one observation to another.
Improved observations
To test the method, the researchers analyzed observations of two stars known to host exoplanets, HIP 65426 and 51 Eridani, along with TYC 5899, a star with no known exoplanets that varies substantially between observations. These are young stars; two had previously been imaged because they host giant planets that remain warm from their gravitational collapse.
Using Webb data and FLIP, the team increased the signal-to-noise ratio by a factor of 1.5 to 2, depending on the observation. For the first time, one of the planets in Webb’s data was detected at a wavelength suitable for identifying carbon dioxide. For TYC 5899, the method increased the sensitivity of the observations by a factor of 15.
The researchers say the technique could do more than provide clearer views of already known planets. Reducing noise could make it possible to find dimmer objects that were previously undetectable, including older or less massive planets. They also believe FLIP could be modified to show where it is making the largest changes to the model, potentially providing useful information about the telescope’s hardware.
Nature, 2026. DOI: 10.1038/s41586-026-11213-z
Source: arstechnica.com
