Exoplanet AI — Cross-Mission Planet Hunter

Kepler 84.7% → TESS 52.6% (domain shift) → 83.2% after retraining. Real NASA cross-mission AI built from Pandhurna.

Comment

Inspiration

NASA's Kepler and TESS telescopes collect data differently. I wondered: can an AI model trained on Kepler work on TESS without changes? This is the exact cross-mission problem NASA faces. From Pandhurna, Madhya Pradesh, I decided to build a planet hunter to test it.

What it does

How we built it

Tech stack: Python, scikit-learn, pandas, Jupyter Notebook, GitHub Pages
Model: RandomForestClassifier (n_estimators=200)
Pipeline:

  1. Cleaned Kepler CSV
  2. Train/test split
  3. Baseline → 84.7%
  4. Direct TESS test → 52.6% (the failure taught us the most)
  5. Scaling + retrain → 83.2%

Challenges we ran into

The biggest shock: the model was excellent on Kepler but failed on TESS. I learned that telescope data distributions differ significantly. Fixed it with feature scaling and mission-specific retraining. Deploying the first GitHub Pages demo was also new for me.

Accomplishments that we're proud of

  • Demonstrated a real NASA cross-mission AI challenge
  • Documented the full journey: 84.7% → 52.6% → 83.2%
  • Open-sourced everything: code, notebook, confusion matrices
  • Representing Pandhurna, MP, India in a global space AI challenge

What we learned

Domain adaptation matters more than raw accuracy in space AI. Understanding why a model fails across missions is the key insight.

What's next for Exoplanet AI

  • Try CNNs on raw light curves
  • Add NASA K2 mission data
  • Build a real-time prediction API

Built with ❤️ from Pandhurna, Madhya Pradesh, India

Built With

Share this project:

Updates