Confusion matrix demonstrating strong classification
Top features ranked by model importance, highlighting the astronomical measurements that most influenced exoplanet classification.

Celesta — Exoplanet Disposition Classifier

Inspiration

Every confirmed exoplanet begins as a tiny dip in a distant star's brightness. NASA's Kepler mission collected thousands of these signals, but distinguishing genuine planets from false positives remains a challenging classification problem.

We wanted to apply machine learning to real astronomical data instead of simplified classroom datasets. This project allowed us to combine data science, artificial intelligence, and astronomy to build a reproducible model capable of classifying Kepler Objects of Interest while emphasising explainability and responsible model development.

What it does

Celesta is a machine learning pipeline that classifies NASA Kepler Objects of Interest into three categories:

✅ Confirmed Planet
🟡 Candidate
❌ False Positive

The project performs data preprocessing, exploratory data analysis, feature engineering, model training, evaluation, and explainability. Rather than simply maximising accuracy, the focus is on building a reliable and interpretable classifier that generalises well to unseen observations.

How we built it

The project follows a complete end-to-end machine learning workflow.

Dataset Preparation

Imported the curated NASA Exoplanet Archive dataset.
Explored feature distributions and missing values.
Removed duplicate and irrelevant attributes.
Cleaned the dataset for reproducible training.

Exploratory Data Analysis

We investigated:

Feature distributions
Class balance
Missing values
Correlations between astronomical measurements

Visualisations helped us understand the structure of the data before training any models.

Preventing Data Leakage

One of the most important parts of the project was identifying target leakage.

Columns including:

koi_pdisposition
kepler_name
koi_comment

contained information that could directly reveal the correct answer. Removing these columns ensured the model learned meaningful astronomical relationships instead of memorizing labels.

Model Development

Multiple classification approaches were evaluated before selecting the final model based on overall performance and interpretability.

The completed workflow includes:

Data preprocessing
Feature engineering
Model training
Cross-validation
Performance evaluation
Confusion matrix
Feature importance analysis

Challenges we ran into

Some of the biggest challenges included:

Handling missing values
Managing class imbalance
Preventing target leakage
Selecting informative features
Avoiding overfitting
Maintaining reproducibility

Working with real scientific data required significantly more preprocessing than typical educational datasets.

Accomplishments that we're proud of

Built a complete machine learning pipeline from raw NASA data.
Successfully detected and removed target leakage.
Produced an interpretable classification model.
Generated meaningful visualisations and evaluation metrics.
Created a reproducible workflow suitable for future improvements.

What we learned

This project strengthened our understanding of:

Data preprocessing
Feature engineering
Exploratory Data Analysis
Classification algorithms
Model evaluation
Explainable AI
Responsible machine learning practices

We also learned that careful data preparation is often more important than choosing increasingly complex algorithms.

What's next for Celesta

Future improvements include:

Hyperparameter optimization
Ensemble learning techniques
SHAP explainability
Probability calibration
Automated feature engineering
Deep learning models
Interactive web dashboard for prediction and visualisation

Ultimately, we hope to extend Celesta into an educational tool that demonstrates how machine learning can assist astronomers in discovering new worlds.