🌟 Exoplanet Detector

TESS light-curve analysis for automatic exoplanet transit detection.

A deep learning pipeline for detecting exoplanet transit candidates in TESS stellar light curves using a hybrid CNN-MLP neural network.

• Deep learning pipeline: see README_DL.md for training & performance
• Full pipeline (data → math → model): see README_PIPELINE.md
• Current DL test accuracy: 60.31% (27 epochs, early stopped from max 80)

---

📋 Overview

This project analyzes TESS (Transiting Exoplanet Survey Satellite) stellar light curves to classify stars as hosting exoplanet candidates or not. The system combines:

• 1D Convolutional Neural Network (CNN) — learns patterns in brightness dips from light curves
• Multi-Layer Perceptron (MLP) — processes Box Least Squares (BLS) transit features
• Hybrid fusion — combines both feature types for improved classification

Classification Task

Label	Meaning	Class
CP, KP, PC	Confirmed / Known / Candidate planet	1 — Planet
FP, FA	False positive / False alarm	0 — Not a planet

---

🎯 Model Performance

Latest Results: 60.31% test accuracy with a hybrid neural network

Metric	Score
Test Accuracy	60.31%
Test F1 Score	52.43%
Test ROC-AUC	63.61%
Best Validation Accuracy	69.5%
Training Epochs	27 (early stopped from max 80)

Dataset Statistics

• Total Samples: 1,598 (after deduplication)
• Planet Candidates: 736
• Non-planets: 862
• Test Set Size: 320 (20% hold-out)
• Train/Val/Test Split: 64% / 16% / 20%

---

🚀 Quick Start

Installation

git clone https://github.com/Ash310u/Exoplanet-detector.git
cd Exoplanet-detector
pip install -r requirements.txt

Train the Model

python3 src/train_cnn.py --epochs 80 --patience 15

Run in Jupyter

Open src/plot.ipynb and run from the "DL MODEL" section.

---

📁 Project Structure

├── README.md                        # Project overview
├── README_DL.md                     # DL training summary & performance
├── README_PIPELINE.md               # Full pipeline: data → math → model layers
├── requirements.txt                 # Python dependencies
├── data/processed/
│   ├── labels_cleaned.csv           # Ground truth labels
│   ├── training_tid_features.csv    # BLS features
│   ├── stars/                       # Light curve files (TIC_*.npy)
│   └── transit_cnn.pt               # Trained model
└── src/
    ├── train_cnn.py                 # Main training script
    ├── plot.ipynb                   # Interactive notebook
    ├── extract_features.py          # BLS feature extraction
    └── build_folded_lightcurves.py  # Phase-folding utility

---

🔬 Pipeline Overview

Light Curves (.npy)    BLS Features (CSV)
        │                       │
        ▼                       ▼
   Preprocess            Normalize
        │                       │
        └───────┬───────────────┘
                ▼
         TransitHybrid Model
         ┌─────────────────┐
         │ CNN Branch      │
         │ BLS Branch      │
         │ Fusion Head     │
         └─────────────────┘
                ▼
         Planet Probability

Model Architecture

CNN Branch: Conv1d → BatchNorm → ReLU → MaxPool (×2) → 32 features

BLS Branch: Linear(5→128) → Linear(128→64) → Linear(64→32)

Fusion: Concatenate + Binary classifier

Training Settings

Setting	Value
Optimizer	AdamW (lr=0.001)
Loss	Binary cross-entropy
Batch Size	32
Early Stopping	15 epochs
Max Epochs	80

---

📊 Performance Comparison

Model	Accuracy
Original CNN	49.4%
Hybrid Model	60.3%
XGBoost Baseline	~63.4%

---

🔧 Improving Accuracy

Light curves are not phase-folded, limiting CNN learning. To improve:

python3 src/build_folded_lightcurves.py
python3 src/train_cnn.py --epochs 80 --patience 15

Phase-folding aligns repeated transit dips, making patterns clearer.

---

📚 Additional Resources

• Full pipeline guide (data extraction → formulas → DL layers): See README_PIPELINE.md
• Detailed DL documentation: See README_DL.md
• Training arguments: Use --help flag with training script

---

💡 Key Technologies

• PyTorch — Neural network framework
• NumPy — Numerical computing
• Pandas — Data manipulation
• Scikit-learn — Preprocessing & metrics
• Jupyter — Interactive notebooks

---

📝 Files Overview

File	Purpose
train_cnn.py	Main training script
plot.ipynb	Interactive notebook
extract_features.py	BLS feature extraction
transit_cnn.pt	Saved model & threshold

---

🎓 Technical Details

Input Features

1. Light curve: 512-point normalized brightness sequence
2. BLS features: Period, depth, duration, SNR, power

Preprocessing

• Clip outliers (5th–95th percentile)
• Detrend with Savitzky–Golay filter
• Robust scale: (x - median) / IQR

Validation Strategy

• Stratified train/val/test split
• Balanced class sampling
• Per-epoch threshold optimization (0.2–0.8)
• Model checkpointing & early stopping

---

🚧 Future Work

• Phase-folded light curves
• Data augmentation (time warping, noise)
• Ensemble methods
• Hyperparameter optimization
• Advanced feature engineering

---

📄 License

Research and educational use.

✨ Quick Reference

Input:    512-point flux + 5 BLS features
Output:   Planet candidate probability
Task:     Binary classification
Model:    TransitHybrid (CNN + MLP)
Accuracy: 60.31% (27 epochs, early stopped from max 80)