Exoplanet Transit Classification with CNNs and Multi View Transformers

This repository contains the code and experiments for our graduate-level CSE project on classifying exoplanet transit candidates using phase–folded light curves from NASA’s TESS mission. We evaluate three architectures — a CNN baseline, a supervised transformer, and two self-supervised extensions — to leverage both labeled and unlabeled astronomical observations.

---

Disclaimer: All collaboration and code execution was done in a Google Colab environment, with the environment containing the necessary files and packages.

---

🚀 Project Overview

Exoplanet detection from light curves is challenging due to:

• high class imbalance (confirmed planets are very rare!)
• stellar variability and instrumental noise
• subtle temporal patterns in transit crossing events (TCEs)

Our approach uses two views of each light curve:

• Global view: full orbital phase (captures context and baseline shape)
• Local view: zoomed window around centroid peak (captures transit structure)

Both flux and centroid signals are encoded as aligned channels.

---

🧠 Model Architectures

Model	Input	Learning Type	Purpose
CNN Baseline	Global flux only	Supervised	Local dip detection
Transformer Encoder	Dual-view flux + centroid	Supervised	Global + local temporal reasoning
Pseudo-Labeled Transformer	Same as above	Semi-supervised	Expand training with confident predictions
SSL Reconstruction Transformer	Same as above	Self-supervised → fine-tuned	Learn transit curve structure without labels

Full methodology details are provided in the project report (report/).

---

📂 Repository Structure

.
├── data/                         # Not included, stored in google drive
│
├── models/
│   ├── cnn.ipynb                 # CNN baseline implementation
│   ├── transformer.ipynb         # Supervised transformer
    ├── pseudo_labeling.ipynb     # Psuedo-labeled transformer
│   ├── ssl_reconstruction.ipynb  # Masked reconstruction transformer (not included in results)
│   └── reading_bulk_data.ipynb   # Utility functions for reading data
│
├── report/
│   └── final_report.pdf
│
└── README.md