Exoplanet Detection and Exploration Platform

Project Overview

Our project is a web-based platform for exoplanet detection and exploration, powered by NASA datasets (Kepler, K2, TESS, KOI). It enables:

1. Prediction of planetary candidates

   • Input via tabular metadata (orbital period, radius, stellar parameters, etc.)
   • Upload of lightcurve data (flux time-series)

2. 3D Visualization of planetary systems

   • Interactive orbital simulations built with Three.js

Quick Start

1. Clone Repository

git clone https://github.com/KienPC1234/ExoVision-Kepler-FPT.git
cd ExoVision-Kepler-FPT

🧪 3. Create and Activate Virtual Environment

python3 -m venv .venv
source .venv/bin/activate  # For Linux/macOS
# Or use .venv\Scripts\activate on Windows

Once activated, your terminal prompt should show (.venv) indicating you're inside the virtual environment.

🐘 4. Install PostgreSQL

On Ubuntu/Debian

sudo apt update
sudo apt install postgresql postgresql-contrib -y

Create database and user

sudo -u postgres psql

-- Inside psql shell:
CREATE USER superuser WITH PASSWORD 'Dukuma6Chi7Bolali';
ALTER USER superuser WITH SUPERUSER;
CREATE DATABASE kepler_app_db OWNER superuser;
\q

🔧 Optional: Customize Database Credentials

If you'd like to change the default database username, password, or connection URL, you can do so in the file:

web/db/base.py

Example:

# web/db/base.py

DATABASE_URL_SYNC = "postgresql+psycopg2://your_username:your_password@localhost:5432/your_database"
DATABASE_URL_ASYNC = "postgresql+asyncpg://your_username:your_password@localhost:5432/your_database"

🔐 Tips:

• Use strong passwords for production environments.
• If you're deploying to cloud (e.g. Railway, Render), replace localhost with your remote host address.
• You can also load these values from a .env file using os.getenv() for better security:

---

🔧 5. Setup Google OAuth2 for ExoVision

To run this app with Google login, create a file at .streamlit/secrets.toml with the following content:

[auth]
redirect_uri = "http://localhost:8501/oauth2callback"
cookie_secret = "Spgg2r4HHGal37TH4uvcsubbfJ_nv3IEHgF7ezLtiBU"

client_id = "YOUR_CLIENT_ID"
client_secret = "YOUR_CLIENT_SECRET"
server_metadata_url = "https://accounts.google.com/.well-known/openid-configuration"

🪪 How to get client_id and client_secret

1. Go to Google Cloud Console
2. Create a new project
3. Enable OAuth consent screen
4. Create OAuth 2.0 Client ID
   • App type: Web
   • Add http://localhost:8501/oauth2callback to Authorized redirect URIs
5. Copy client_id and client_secret into the file above

---

6. Install Requirements

pip install -r requirements.txt

7. Run Web App

streamlit run streamlit_app.py

---

How It Works

• Tabular Metadata → Ensemble Stacking (LightGBM, Random Forest, XGBoost, Neural Network) with Logistic Regression meta-learner.
• Lightcurves → Transformer model (PatchTST) with attention on patched windows.
• Preprocessing Pipeline includes dataset merging, cleaning, unit conversions, label encoding, feature engineering, imputation, scaling, balancing (SMOTE), and artifact saving for consistent inference.

Predictions are logged and displayed in both tabular and 3D orbital views. Users can also train custom models with built-in tutorials.

---

Project Structure

ExoVision/
├── .streamlit/               # Streamlit configuration files
│   └── config.toml
├── LICENSE                   # Project license information
├── ModelTrainer/             # Core training and model loading logic
│   ├── checkgpu.py           # GPU availability checker
│   ├── modelV1/              # Version 1 of the model pipeline
│   │   ├── data_preprocess.py
│   │   ├── model_builder.py
│   │   ├── model_loader.py
│   ├── modelV2/              # Version 2 of the model pipeline
│   │   ├── check_dataset.py
│   │   ├── data_preprocess.py
│   │   ├── model_builder.py
│   │   ├── model_loader.py
│   └── readme.md             # Internal documentation for ModelTrainer
├── data/                     # Preprocessed and raw data files
│   ├── koi_lightcurves.parquet
│   ├── merged_processed.csv
├── dataset/                  # External datasets used for training and prediction
│   ├── k2_pandc.csv
│   ├── koi_cumulative.csv
│   ├── toi.csv
├── models/                   # Saved models and evaluation artifacts
│   ├── v1/                   # Artifacts from model version 1
│   │   ├── feature_list.pkl, stacking_model.pkl, etc.
│   ├── v2/                   # Artifacts from model version 2
│   │   ├── best_patchtst.pth, y_test.npy
├── readme.md                 # Main project documentation
├── requirements.txt          # Python dependencies
├── static/                   # Static assets for Streamlit (fonts, videos, etc.)
│   ├── *.ttf, *.mp4
├── streamlit_app.py          # Entry point for the Streamlit web application
├── supervisor_config/        # Supervisor configuration files for deployment
│   ├── iframe_loader.conf, streamlit.conf
├── usgi_service.py           # uWSGI service integration script
├── web/                      # Backend and frontend logic for web integration
│   ├── db/                   # Database models and wrappers
│   ├── helper/               # Custom Streamlit components and helpers
│   ├── iframe_loader/        # HTML/JS assets for iframe rendering
│   ├── pages/                # Streamlit page modules (e.g., home, login, prediction)
│   ├── utils/                # Utility functions (auth, routing, etc.)

---

🎬 Kepler Project Showcase

Explore our latest AI-driven research on exoplanet detection and data visualization. Below are two short demo videos showcasing ExoVision’s data processing and prediction systems in action.

---

🚀 Demo 1 — Exoplanet Predictor

KeplerFPTC.mp4

---

🔭 Demo 2 — Exoplanet Flux Predictor

KeplerFPT2C.mp4

---

🧠 Kepler-FPT combines deep learning with astrophysical insights to uncover hidden worlds beyond our solar system.

Benefits

• High Recall & Accuracy (~90% for metadata, >85% for lightcurves) → fewer missed candidates.
• User-Friendly: Simple UI for predictions, uploads, search/export history.
• Interactive: Realistic 3D orbital viewer for deeper insights.
• Efficient & Secure: Async worker queue, GPU acceleration, authentication, HTTPS.
• Extensible: Support for custom model training and experimentation.

---

Intended Impact

The platform democratizes exoplanet research by lowering barriers to entry:

• Researchers → rapid prototyping and exploration.
• Educators → classroom-friendly simulations.
• Students & citizen scientists → accessible tools for discovery.

Our focus on high recall addresses the "needle-in-a-haystack" challenge, ensuring rare planetary candidates are not overlooked.

---

Tech Stack

• Languages: Python (ML backend), JavaScript (Three.js frontend)
• Frameworks: Streamlit (UI), PyTorch (PatchTST), scikit-learn, LightGBM, XGBoost
• Data Tools: Pandas, NumPy, PyArrow, Astropy
• Deployment: Docker, Celery/RabbitMQ, Nginx, GPU VPS
• Other Tools: Git, Jupyter, Pickle/Joblib

---

Creativity & Innovation

• Combines ensemble tree models (tabular) with transformer models (time-series) in one platform.
• Real-time 3D visualization brings planetary systems to life.
• User-oriented workflows (prediction history, custom training) make research collaborative.
• PatchTST patching handles noisy flux data with minimal feature engineering.

---

Design Considerations

• Scalability: Efficient GPU-based processing for billions of flux points.
• Usability: Streamlit UI for non-experts, tutorials for new users.
• Ethics & Security: Bias mitigation (SMOTE), authentication, privacy compliance.
• Performance Trade-off: Prioritized recall over precision.
• Team Collaboration: Modular code for parallel development.

---

Future Directions

• Real-time TESS data streaming for live candidate detection.
• Expanded citizen science features for open participation.
• Multi-class classification expansion with more nuanced planetary states.

---

🚀 Making exoplanet research accessible, interactive, and collaborative for everyone.