Improving the Search for Exoplanets with AI
If you know what to look for, when you gaze out into the night sky you can probably find a few of the other planets in our solar system amongst the vast array of stars. It’s difficult to see planets beyond our solar system even with powerful telescopes—they are small and dim and obscured by the overwhelming light of the star they orbit. To find exoplanets, astronomers have therefore devised clever methods to detect them indirectly through the influence they have on their stars.
“We have two methods that have been highly prolific at finding exoplanets,” said Eric Ford, distinguished professor of astronomy and astrophysics and a co-hire of the Penn State Institute for Computational and Data Sciences. “The first, called the transit method, looks for a dimming of a star’s light when an orbiting planet passes between the star and an observatory. The second looks for shifts in the spectra of light from a star that wobbles due to the gravitational effect of orbiting planets.”
Ford, who is director of the Penn State Center for Exoplanets and Habitable Worlds, studies planet formation and the evolution of planetary systems. He said that for this second method—known as the radial velocity, or Doppler, method—one of the challenges is distinguishing signal from noise.
“Our instruments, like the NEID and Habitable-zone Planet Finder—both built here at Penn State—are now so good that the main source of noise is no longer from measurement error,” said Ford. “Instead, the noise comes from the fact that stars are not ideal billiard balls emitting light. Stars are convective, pulsating balls of hot gas with dark spots where there are strong magnetic fields. The magnetic activity imprints on the spectra of light we see from the star, potentially muddling the signal of a planet. Observing a star very often could enable us to account for this variability, but there is limited observing time. For most stars we don’t have the 100s or 1000s of observations that might be required to understand how stellar variability affects that star’s spectrum.”
To help overcome this barrier, Ford and his research team are turning to AI and the one star for which we do have enough observations, our own sun.
“We have hundreds of thousands of spectra of the sun, so we can use that data as a training set for machine learning models. The first step is to learn how to extract accurate velocity measurements for the Sun. The next step will be to adapt those methods so they can be applied to other stars,” said Ford. “We are using a hybrid AI approach, what I call ‘physics-informed machine learning.’ For example, we understand the physics of the Doppler effect very well, so we can feed that into the model rather than forcing the model to ‘learn’ that on its own.”
Ford hopes the model can help inform future studies by reducing the number of observations needed to account for the variability in spectra caused by stellar activity.
“If you want to train AI to tell the difference between a cat and a dog, it can pull data from millions of photos on the internet. The training data needed is cheap and readily available,” said Ford. “This is not the case for data obtained by state-of-the-art instruments in exoplanets and many other scientific fields. With the solar data, we can ask: ‘What if we only had a selection of 100 or 1000 of these observations?’ How much data do we really need to be able to discover rocky planets around other stars?”
While AI is often employed to tackle problems of too much data, Ford explained that the science-informed approach to AI can be particularly fruitful for fields like finding exoplanets, where algorithms need to be designed to make the most of limited specialized data that is much harder to come by.
Editor's Note: This story is part of a larger feature about artificial intelligence developed for the Winter 2026 issue of the Eberly College of Science Science Journal.