🌌 A World Away — Exoplanet Classifier

NASA Space Apps Challenge 2025

Overview

A World Away is an artificial intelligence tool for detecting and classifying exoplanets using open data from NASA’s Kepler, K2, and TESS missions.

The app enables users to:

• Train an AI model using real NASA data.
• Predict whether an object is a confirmed exoplanet, a candidate, or a false positive.
• Explore how physical parameters (radius, transit duration, stellar temperature, etc.) affect the classification.
• Interact with predictions through an educational, web-based interface.

Objective

To accelerate and democratize exoplanet detection through interpretable machine learning, making it accessible both to researchers and enthusiasts.

Data Sources

• Kepler Cumulative Table
• TESS Object of Interest Table
• K2 Candidate Catalog

Research references:

• MNRAS 2021: Machine Learning for Exoplanet Classification
• Electronics 2024: ML Techniques for Exoplanet Discovery

Model pipeline

1. Preprocessing

• Loads NASA’s Kepler dataset and ignores commented header lines.
• Cleans missing or invalid values and normalizes key numeric features.
• Filters to the subset of physically meaningful variables.

2. Feature selection Each chosen feature corresponds directly to a measurable property in the transit method:
   • koi_period — orbital period (days)
   • koi_duration — transit duration (hours)
   • koi_depth — transit depth (ppm)
   • koi_prad — planet radius (Earth radii)
   • koi_srad — stellar radius (Solar radii)
   • koi_steff — stellar effective temperature (K)
   • koi_slogg — stellar surface gravity (log g)
3. Model Benchmarking & Training

• Trains and compares RandomForest, XGBoost, and CatBoost classifiers on the labeled classes: CONFIRMED, CANDIDATE, and FALSE POSITIVE
• Selects the model with the best accuracy–interpretability trade-off based on validation metrics (Accuracy, F1, ROC-AUC).

4. Export

• Saves the best-performing model and the fitted label encoder in /models/.
• Generates model_results.json and benchmark plots for documentation.

5. Class Means

• Stores per-class mean values for each feature, used to produce interpretative comparisons (e.g., how a candidate differs from typical confirmed planets).

Web Application

Built with Streamlit, the app allows:

• CSV upload for batch predictions.
• Manual input for single-candidate prediction.
• Pie chart of predicted class probabilities.
• Educational interpretation comparing user inputs with per-class averages.
• Global and local SHAP visualizations for explainability.

Technical Stack

Component	Technology
Backend / Model	Python 3.10, scikit-learn, pandas, joblib
Frontend	Streamlit + Plotly
Explainability	SHAP (planned)
Styling	Custom CSS

Installation & Run

git clone https://github.com/arnoop88/ExoAI.git
cd ExoAI
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
python3 src/train_model.py
python3 src/benchmark_models.py
python3 src/explain_model.py
streamlit run src/app.py