NASA Space Apps Challenge 2025

Exoplanet AI

Cross-Mission Planet Hunter

Trained on Kepler. Tested on TESS. Solved domain shift.

View on GitHub How it works
Kepler Train
84.7%
Baseline accuracy
TESS Zero-Shot Fail
52.6%
Domain shift hits
TESS Adapted Win
83.2%
After scaling
Kepler → TESS transfer learning

The Problem

Domain shift between Kepler and TESS

Same universe, different cameras. Kepler stared deep at one patch for 4 years. TESS scans the whole sky in 27-day sectors. Different noise, cadence, and stars = an AI trained on Kepler gets confused on TESS.

Kepler Mission

  • • Fixed 115 sq-deg field
  • • 30-min long cadence, 4 years
  • • Faint, quiet stars
  • • Low noise, stable

TESS Mission

  • • All-sky survey, 26 sectors
  • • 2-min & 30-min cadence
  • • Brighter, nearer stars
  • • Higher systematics
!

What is domain shift?

Think of it as accent difference. The model learns Kepler's "accent" — its noise patterns and timing. When we show it TESS data, the words (transits) are the same, but the accent is different. Without adaptation, accuracy crashes from 84.7% → 52.6%.

Results

From failure to 83.2%

Placeholders for your actual plots — export from notebook as PNG and swap the boxes.

Kepler Confusion Matrix

84.7% accuracy

Train → Test on Kepler Strong

TESS Real Confusion Matrix

52.6% zero-shot

Kepler model → TESS Domain shift

ROC Curve

Adapted 83.2%

After scaling + retrain Recovered

How it works

4 steps to cross-mission hunting

STEP 01

Load Kepler data

NASA Exoplanet Archive cumulative KOI table. Clean features: period, duration, depth, SNR, stellar temp.

1
STEP 02

Train RandomForest

80/20 split, balanced classes. Achieves 84.7% on Kepler holdout.

2
STEP 03

Test on TESS (fail)

Direct transfer drops to 52.6%. Model overfits to Kepler noise fingerprint.

3
STEP 04

Scale + Retrain (83.2%)

StandardScaler on combined features, retrain on Kepler, evaluate on TESS. Domain gap closes.

4

Tech Stack

Python • scikit-learn • pandas • NASA data

Python 3.11 scikit-learn pandas matplotlib NASA Exoplanet Archive
Nohar, Rajasthan, India

Built for NASA Space Apps Challenge 2025 by Parineeta Pahade

Open-source cross-mission exoplanet detection. From Kepler's deep stare to TESS's all-sky sweep — one model, adapted.

Star the repo