Welcome to StellarNet

The Tool That Helps Amateur Astronomers to Discover Exoplanets. Trained on profiessional and Amateur data.

Testing Loss	0.0811
Accuracy	0.9794
F1-score	0.6000

Confusion Matrix (Test Data)

	Pred: No	Pred: Yes
Actual: No	534	1
Actual: Yes	9	9

Model Dashboard

Architecture

Our exoplanet detection system uses a 1D Convolutional Neural Network (CNN) specifically designed to analyze time-series light flux data.

Network Structure

Convolutional Layers:
• Conv1D (1 → 32 channels, kernel=5) + BatchNorm
• Conv1D (32 → 64 channels, kernel=5) + BatchNorm
• Conv1D (64 → 128 channels, kernel=3) + BatchNorm
• MaxPooling (stride=2)
• Dropout (0.3)

Fully Connected Layers:
• FC1 (flattened → 128 units)
• FC2 (128 → 2 classes)

Training Data

The model was trained on a diverse dataset combining:

Lightkurve API data - Professional light flux measurements from NASA's Kepler missions
Amateur observations - Community-contributed data to improve real-world applicability

Data Processing

All input data is scaled and standardized to ensure consistent predictions regardless of the source telescope or observation conditions. We automatically scale and transform data submitted by users.

Handling Class Imbalance

Since exoplanets are rare, we employed:

Class weighting - Increased importance of positive exoplanet examples during training
Data augmentation - Generated additional synthetic exoplanet signals to balance the dataset