Whose light is dipping?
If the dimming actually comes from a neighboring star, it’s not your planet. This clue — the centroid offset — catches more fakes than any other.
Interactive · Real NASA data · Machine learning
Finding Worlds
in the Noise
A machine-learning model that reads the flicker of distant stars — and tells real planets from cosmic impostors.
Drag the planet to spin it
The problem
Most “planets” aren’t planets.
When a world crosses in front of its star, the starlight dims — a tiny, telltale dip. But two overlapping stars can fake that dip perfectly. So can plain instrument noise. Separating real worlds from impostors is slow, careful, expert work.
Three nearly identical flickers. Only one is a world.
The data
Nine and a half thousand whispers,
caught by a real telescope.
Every signal here is real. It comes from NASA’s Exoplanet Archive — the Kepler Objects of Interest catalog — recorded by the Kepler Space Telescope as it stared at a single patch of sky, watching thousands of stars for the faintest flicker.
~0real observed signals
Dozensof measurements per signal
0+stars watched by Kepler
Source · NASA Exoplanet Archive
Catalog · Kepler Objects of Interest (cumulative)
Mission · Kepler Space Telescope
Confirmed planet
False positive
Candidate
Every point of light is one real Kepler signal — 9,500 in all. Drag to explore.
The model
Three clues. One verdict.
Two specialists — a Random Forest and an XGBoost model — each study every signal. A third “manager” model weighs their opinions and makes the final call. The teamwork is called a stacking ensemble. Here’s what it looks for:
Random ForestXGBoost→Manager model→Verdict
How big is the shadow?
Planets are small. If the crossing object looks too large, it’s almost certainly another star — not a world.
How clean is the signal?
Strong, crisp, repeating dips earn trust. Faint, messy ones don’t. Astronomers call it signal-to-noise.
The same checks human astronomers use to vet planets — the model just never sleeps.
0%accuracyacross ~9,500 real signals
0.00Macro F1balanced score across all three verdicts
0.00cross-validatedholds up on data it never saw
Try it — make a transit
brightnesstime →
Move your cursor (or finger) across the panel — watch the starlight dip as the world crosses its star. This dip is exactly what the model reads.
The verdicts
Every signal gets one of three calls.
A real exoplanet
The signal passed every test. Somewhere out there, a world is circling its star.
An impostor
It looked like a planet, but isn’t — here, two eclipsing stars faking the dip. Or just noise in a planet costume.
Not sure — yet
The genuinely ambiguous pile: promising, flagged, and still awaiting a final answer.
Why it matters
Faster vetting. Honest science.
More worlds.
Kepler left thousands of signals awaiting judgment — and new telescopes add more every night. A model that handles the first pass of vetting, accurately and honestly, frees scientists to spend their time where it counts: confirming real worlds.
placeholder links — search “TODO-LINKS” in the code to drop in your URLs