Project Story
Inspiration
Space exploration has always generated enormous amounts of scientific data, but analyzing that data manually is both time-consuming and resource-intensive. Inspired by the challenge of applying artificial intelligence to a real NASA dataset, we wanted to build a machine learning solution that could help identify potential exoplanets more efficiently. The opportunity to work with authentic Kepler mission data while solving a meaningful scientific problem made this project especially exciting.
What it Does
Our project uses machine learning to classify Kepler Objects of Interest (KOIs) into one of three categories:
- Confirmed – Verified exoplanets.
- Candidate – Promising signals requiring further investigation.
- False Positive – Signals caused by noise or other astronomical phenomena.
By analyzing transit characteristics, stellar properties, and orbital parameters, the model predicts the likelihood that a detected signal represents a real exoplanet. This helps automate one of the most important stages of exoplanet discovery and demonstrates how AI can accelerate scientific research.
How We Built It
We started by exploring the NASA Kepler dataset, which contains thousands of observations and over a hundred astrophysical features. The first step involved understanding the data distribution, identifying missing values, and preparing the dataset for training.
After preprocessing, we selected relevant features describing planetary and stellar characteristics, handled missing data, and encoded the target classes. We then trained and evaluated several machine learning models, comparing their performance using metrics such as Accuracy, Precision, Recall, and F1-score.
Feature importance analysis helped us understand which astronomical measurements contributed most to the model's predictions, making the system more interpretable alongside its predictive performance.
Conceptually, the model learns a function
$$ f(X) \rightarrow \text{Planet Classification} $$
where (X) represents the observed stellar and transit measurements, and the output is the predicted disposition of each candidate.
Challenges We Faced
One of the biggest technical challenges was working with a large scientific dataset containing numerous features and missing values. Determining which attributes were most informative required careful preprocessing, feature engineering, and experimentation with different machine learning approaches. We also had to address class imbalance while ensuring the model generalized well across all three classes rather than favoring the majority class.
In addition to the technical hurdles, we encountered an unexpected hardware issue during development. My laptop stopped working partway through the hackathon, which meant I had to continue the project using my brother's laptop and GitHub account. As a result, the GitHub commit history reflects activity from his account rather than mine. Despite this setback, we adapted quickly, continued collaborating, and successfully completed the project within the hackathon timeframe. The experience reinforced the importance of resilience, adaptability, and problem-solving under pressure.
What We Learned
This project strengthened our understanding of the complete machine learning workflow—from exploratory data analysis and feature engineering to model evaluation and interpretation.
More importantly, we learned how AI can contribute to real scientific discovery. Building a classifier for NASA's Kepler observations demonstrated that machine learning is not only useful for commercial applications but can also assist researchers in accelerating discoveries in astronomy and space science.
Accomplishments That We're Proud Of
- Successfully built an end-to-end machine learning pipeline for exoplanet classification.
- Worked with a real-world NASA scientific dataset containing thousands of observations.
- Improved our skills in data preprocessing, feature engineering, model evaluation, and explainable AI.
- Demonstrated how machine learning can support astronomical research by identifying promising exoplanet candidates efficiently.
What's Next
Future improvements include experimenting with ensemble and deep learning models, incorporating additional astronomical datasets, improving model explainability using SHAP or similar techniques, and developing an interactive web application that allows users to upload new observations and receive predictions in real time. Ultimately, we envision creating a tool that makes AI-assisted exoplanet discovery more accessible to students, researchers, and astronomy enthusiasts.
Log in or sign up for Devpost to join the conversation.