ExoDetect AI

Inspiration

Astronomers rarely observe exoplanets directly. Instead, they analyze subtle changes in starlight to infer the presence of planets orbiting distant stars. We were inspired by the idea that the same machine learning techniques used in industry can also contribute to real scientific discovery. This challenge provided an opportunity to apply AI to authentic NASA data while learning how data science supports modern astronomy.

What it does

ExoDetect AI is a machine learning system that classifies exoplanet candidates using observational data from the NASA Exoplanet Archive. It analyzes astronomical features, performs data preprocessing, and predicts whether a candidate is a genuine exoplanet or a false positive. Along with accurate predictions, the project focuses on explainability through feature importance analysis and visualizations.

How we built it

We built ExoDetect AI using Python and the scientific machine learning ecosystem.

Our workflow included:

Exploratory Data Analysis (EDA)
Data cleaning and preprocessing
Missing value handling
Feature engineering
Class imbalance handling
Model training and hyperparameter tuning
Performance evaluation using Accuracy, Precision, Recall, F1-score, ROC-AUC, and Confusion Matrix
Explainability using feature importance and model interpretation techniques

The project was developed in Jupyter Notebook using libraries such as Pandas, NumPy, Scikit-learn, Matplotlib, and XGBoost/LightGBM (where applicable).

Challenges we ran into

Working with real-world scientific data introduced several challenges. The dataset contained missing values, correlated features, and class imbalance that required careful preprocessing. Choosing evaluation metrics beyond simple accuracy was important because predicting rare exoplanets correctly is more valuable than achieving a high overall accuracy. Balancing model performance with interpretability was another key challenge.

Accomplishments that we're proud of

Built a complete end-to-end machine learning pipeline.
Worked with a real NASA dataset instead of synthetic data.
Produced an explainable model rather than treating it as a black box.
Generated meaningful visualizations to understand feature relationships.
Created a reproducible workflow suitable for future research and improvements.

What we learned

This project strengthened our understanding of data preprocessing, feature engineering, classification algorithms, and model evaluation. More importantly, we learned how machine learning can support scientific research by extracting meaningful insights from complex observational data. We also gained experience in building reproducible and interpretable AI systems.

What's next for ExoDetect AI

Future improvements include testing deep learning approaches, incorporating additional NASA datasets, improving feature engineering using astrophysical knowledge, experimenting with ensemble methods, and developing a web application where users can upload observational data and receive explainable predictions. We also plan to integrate SHAP/LIME explanations to provide greater transparency into model decisions.