AI COMMANDERTHE EXOPLANET HUNTER
Design a space telescope, launch it, and hunt for exoplanet signals in real NASA Kepler data, with an AI Scientist as your second opinion.
Play live: ai-commander-nasa.vercel.app
Not affiliated with or endorsed by NASA. Uses historical NASA Kepler data; engineering values are educational approximations.
Mission design is hard to learn from a textbook
Students read about space missions. Engineers have to decide. That gap is where beginners lose interest.
Every choice is a trade-off
A bigger telescope means more mass, more power and more cost. Beginners read this, but rarely feel it.
Systems depend on each other
Pick a distant orbit and you also need a stronger rocket and a deep-space ground network.
Real science data feels out of reach
Exoplanet hunting means reading faint brightness dips that look like noise to a newcomer.
One Mission. Six Critical Decisions.
From spacecraft design to exoplanet discovery.
Every choice is a real engineering trade-off
Each part changes several resources at once. The backend mission engine recalculates everything live.
Every part costs money; go over and you cannot launch.
The rocket must lift it to the chosen orbit.
Spare power becomes energy for observations.
All science data must reach Earth.
Risky parts and thin margins raise the chance of launch problems.
Launch stays blocked until every critical check passes: budget, rocket capacity, power, communications, ground station, lifetime, energy and risk.
These are the same kinds of trade-offs real mission teams study, simplified for learning. Values are educational approximations, not NASA engineering data.
Real Kepler data, with an AI second opinion
When a planet crosses in front of its star, it blocks a tiny amount of light. Scientists look for these brightness dips.
A processed, phase-folded Kepler light curve from the game (128 points, ±3 transit durations).
- 1Kepler photometry from MAST
- 2Phase-folded on the catalog period
- 3128 median bins, ±3 transit durations, in ppt
- 4Checked against Kepler DR25 KOI labels on the server
- Model
- 1D convolutional neural network (TensorFlow), 166,849 parameters
- Data
- 1,000 Kepler DR25 KOIs (500 candidates / 500 false positives), split by star 700 / 150 / 150
- Held-out test (150 stars)
- Accuracy 71.33% · candidate recall 94.67% · threshold 0.40
- Runs live in the backend; no mock scores
- A raw experimental score, never a calibrated probability
- Flags many false positives as candidates; it can be wrong
- Never confirms a planet: the player makes the call
Only the 150 test-split stars are playable, so these metrics are not an independent new benchmark.
Built on open tools and NASA resources
Learn by doing
Players experience the mission life cycle: objective, requirements, design, verification, launch, operations and science.
Honest about AI
Players see that AI helps, but can be wrong, and learn to weigh it against evidence.
Open to everyone
Free in the browser, works on phones, with guided play for beginners.
- NASA — Design and Develop Science Missions (PDF)Mission life cycle and the four mission segments
https://www.nasa.gov/wp-content/uploads/2025/04/250323-design-and-develop-science-mission.pdf - NASA Goddard Mission Design LabConcurrent engineering trade studies
https://etd.gsfc.nasa.gov/capabilities/integrated-design-center/mission-design-lab/ - NASA Webb "Build It Yourself" gameBeginner-friendly design flow
https://science.nasa.gov/mission/webb/game-build-it-yourself/ - NASA Exoplanet Archive — Kepler DR25 KOI tableCANDIDATE / FALSE POSITIVE labels
https://exoplanetarchive.ipac.caltech.edu/ - MAST — Kepler mission dataSource light curves
https://archive.stsci.edu/missions-and-data/kepler
Not affiliated with or endorsed by NASA. Uses historical NASA Kepler data; engineering values are educational approximations.
Live mission demo
This is the real game, running live. Take command.
- 1Tap START MY MISSION
- 2Choose Guided Explorer
- 3Auto-Build, check and launch
- 4Find the dip and make your call
Click inside the phone to play. Click outside it to use slide controls again.