Open-Source AI Exoplanet Tools Gear Up for NASA's Roman Data

What happened

NASA's Nancy Grace Roman Space Telescope is moving from launch toward science operations, and the software that will sift its data is being built in the open. In early October 2026, Roman demonstrated very stable pointing and gave its planet-imaging coronagraph a first look at the sky 5. Weeks earlier, the observatory had departed on its roughly million-mile trip to the L2 point, where it will survey large swaths of the cosmos for evidence about dark matter, dark energy, and distant planets 5.

In parallel, NASA's AI planet-finder ExoMiner++ has been released as open-source software and is now working through data from the Transiting Exoplanet Survey Satellite (TESS) 23. On its first pass, it flagged about 7,000 TESS signals as exoplanet candidates, which are likely planets that still need confirmation from other telescopes 23. The tool can be downloaded freely from GitHub 2.

Two kinds of open tools, two ways of finding planets

Roman will hunt planets in two very different ways, and each one has its own open-source toolchain.

The first is transits, the small dips in brightness when a planet crosses its star. ExoMiner++ is built for this work. Its earlier version validated 370 exoplanets, and Open Source For You describes that track record as proof it can separate real planetary signals from false positives 3. Phys.org reports that Roman is expected to record tens of thousands of transits. Those data will be public, as TESS data are, and the advances behind ExoMiner could carry over to Roman's archive 2.

The second is direct imaging through the coronagraph, which blocks starlight so faint companions become visible. Northwestern University's Jason Wang has proposed joining the Roman Coronagraph team to adapt two community tools he leads [1]:

• pyKLIP removes the star's glare (stellar point-spread function subtraction). It is already widely used on coronagraphic data from ground-based observatories and JWST 1.
• orbitize! fits planetary orbits using Bayesian statistics. It could predict where known imaged planets, and planets found indirectly through Gaia astrometry, will sit relative to the coronagraph's field of view. That would show which targets need ground-based follow-up to refine their orbits first 1.

The proposal covers target selection, data simulation and reduction, and characterizing how well the instrument performs 1.

Where the accounts diverge

The reports agree on the main numbers: 370 validated planets and roughly 7,000 TESS candidates. They differ in emphasis. Phys.org treats ExoMiner++ partly as preparation for Roman 2. Open Source For You frames the release mainly as a step toward transparent, collaborative science 3. Wang's proposal is a plan, not a finished result, so whether pyKLIP and orbitize! become part of Roman's official pipeline depends on that work going forward 1.

The navigation side story, and a security angle

Another NASA result from the same period points to a related trend. The FALCON experiment on the four-satellite Starling mission showed that spacecraft can work out their position without GPS. It used existing star-tracker cameras to treat other spacecraft and orbital debris as reference points 45. ScienceDaily reports that within three days FALCON also improved the known orbits of more than 200 space objects 5. The timing differs between the two reports. ScienceDaily dated its coverage to late August 2026 5, while Born To Engineer says the results were released in early September 4.

Neither report presents FALCON as a cybersecurity measure. Still, navigation that does not depend on an external radio signal is relevant to anyone worried about GPS jamming or spoofing. That connection is my reading, not a claim NASA made. Better tracking of debris also helps with keeping orbits safe as they get more crowded.

Why it matters

Roman's raw data will be public no matter what. The bigger story is that NASA and its partners are also making the analysis tools public. Open code means any researcher can rerun ExoMiner++ on the same TESS light curves and check, challenge, or extend its candidate list 2. For a model whose output is thousands of unconfirmed signals, that kind of outside checking matters.

There are tradeoffs. A list of 7,000 candidates is a heavy follow-up load, and every candidate still needs confirmation from other telescopes 23. Open repositories also need ongoing maintenance and review of contributions. Even so, the overall direction is sensible: public data, public algorithms, and community tools already proven on JWST and ground-based instruments 1. Roman's large survey output will probably be more useful because of it.