Stardust — finding asteroids in old sky photos

Independent research · with Jossaya Camille · 2026–

When a near-Earth asteroid is discovered, astronomers have only a few days of sightings, so its orbit is fuzzy and its odds of hitting Earth are fuzzy with it. In early 2025, 2024 YR4 briefly carried about a 3% chance of impact in 2032, and it took roughly two months of new observations to rule that out. Ruling it out required more sightings, and the usual way to get them is to wait and keep looking.

There's a faster way. Public sky surveys photograph the whole sky every few nights, so a newly found asteroid was usually already sitting in older photos that nobody had connected to it. Stardust goes back and finds it there, automatically.

How it works

Early orbit. Fit an orbit to only the sightings that existed four days after discovery.
Clones on the GPU. Sample tens of thousands of orbits consistent with those sightings and run them back in time with our own batched N-body integrator in PyTorch: the Sun, planets, Moon, and the most massive asteroids, plus the Sun's relativistic correction.
Search boxes. Project that cloud of clones onto every older photo from the Zwicky Transient Facility (ZTF) at Palomar, which tells us where in each frame the asteroid must have been.
Detect. Search those regions with a U-Net trained on real ZTF cutouts of known asteroids, alongside a classic streak matched filter.
Link and check. Keep a track only if it lines up on a single orbit across several photos and nights, and still fits the real discovery sightings once the orbit is refit with it.

Results so far

We ran it on 72 asteroids that NASA had once listed as possible impactors, giving the pipeline only what was known four days after each discovery. It recovered 167 verified ZTF sightings of 22 of them, none of which had ever been attributed to those asteroids. Adding them makes the orbits up to 52× more certain, with a median of 7.1×. Sixteen of the 22 came from held-out rounds that we ran once, with every rule frozen beforehand.

The AI detector is where the extra reach comes from. At the camera's detection limit it finds 58% of real asteroids, where the classic streak filter finds none. It also stays honest. It doesn't report asteroids that are too faint to be in the photo.

We also pointed it at asteroids on NASA's Sentry risk list today and turned up 202 pre-discovery ZTF sightings of 23 of them that had never been reported, each found independently by both detectors.

To check that the engine underneath is right, we compared it with NASA JPL on 2024 YR4. Our fitted orbit lands within 6 km of JPL Horizons, the propagator stays within 4 km per year of REBOUND+ASSIST, and the predicted 2032 Earth miss distance comes out at 279,443 km against JPL's 278,333 km.

Who did what

I built the orbit and detection side: the GPU propagator, the search and linking, the AI Spotter and its training data, and the checks against JPL. Jossaya built the live site, a 3D dashboard that flies you to each asteroid and shows where Stardust saw it before anyone else did.

Where it's going

The work now is making every result rigorous enough to stand as research, and getting the recovered sightings to the Minor Planet Center.

Numbers

167: Verified, never-connected ZTF sightings of 22 once-risky asteroids.
52×: Largest tightening of an orbit (2025 YH5). Median 7.1×.
58% vs 0%: AI Spotter against the classic streak filter, at the camera's limit.
23 s: 100,000 orbit clones run a year forward, on a laptop GPU.
202: Unreported pre-discovery sightings of asteroids on NASA's risk list today.

Python and PyTorch (CUDA), Find_Orb, REBOUND and ASSIST, Astropy; ZTF data via IRSA/IPAC, the Minor Planet Center, and NASA JPL. Started in October 2026 at Knight Hacks IX. The model, dataset and results page are on Hugging Face. See also PolyOculus and CELERIS.