ISRO BAH 2026

AI Exoplanet Transit Detection Pipeline

The Escapists — Bharatiya Antariksh Hackathon 2026

---

An AI-based pipeline and dashboard for automatically detecting exoplanet transit signals in noisy astronomical light-curve data. Transit photometry hunts for the tiny, periodic dips in a star's brightness caused by a planet crossing its disk. In crowded fields these signals are buried under detector noise and contamination from blended foreground/background sources, and they are easily confused with eclipsing binaries or starspots. This project pairs a deep-learning classifier with an interactive mission-control dashboard to surface and explain candidate transits.

---

Features

📊 Interactive Dashboard	Visualises the full detection workflow across five panels: raw light curve, AI pipeline diagram, denoised signal, transit probability over time, and a phase-folded detection summary.
🔭 Three Input Modes	Feed the model a phase-folded image, a time,flux CSV, or a TESS TIC ID. CSV/TIC run the full pipeline server-side and repopulate every panel from the real data.
📦 Batch & Export	Submit many inputs at once (multiple files or a list of TIC IDs); they process sequentially with a live progress bar, and all results download together as one JSON file.
🤖 Live Model Inference	The trained ResNet18 classifier returns a verdict, confidence score, and per-class probability bars — updated in real time.
🧭 Guided Onboarding	A first-visit welcome and an "Analyze Your Data" call-to-action point newcomers to the input panel; the default view is a clearly-labelled sample, not stored data.
🔍 Help & Detail Views	Every section has a ? explainer and a fullscreen ⤢ expand that opens an in-depth reference panel (and a detailed pipeline flowchart).
🎨 Design System	Single-screen dark layout, sharp corners, greyscale + electric blue (#00e8f7), subtle animations, and an icon splash/loading screen. Documented in DESIGN.md.

---

Tech Stack

Layer	Technology
Backend	Python · Flask · gunicorn (production)
ML / Inference	PyTorch · torchvision (ResNet18) · Pillow
Frontend	HTML · CSS · Vanilla JavaScript · Chart.js
Analysis & Training	Lightkurve · Astroquery · matplotlib · NumPy
Fonts	Bitcount Prop Single (headings) · Ubuntu (body)

---

The Model

The classifier is a ResNet18 fine-tuned on 224×224 RGB images of phase-folded light curves — black scatter on a white background — normalised with standard ImageNet statistics. The full training process is documented in model/tessnet.ipynb.

Raw TESS Data  →  BLS Period Search  →  Phase-fold  →  Render Image  →  ResNet18 Fine-tune

1. Download TESS light curves via Lightkurve (SPOC pipeline).
2. Clean — remove NaNs, sigma-clip outliers, flatten the trend.
3. BLS periodogram — find the dominant period, epoch, and duration.
4. Phase-fold the light curve and render it as a 224×224 PNG.
5. Fine-tune a pre-trained ResNet18 with a custom classification head.

Trained weights ship as frontend/tess_resnet18_model.pth. Inference mirrors the notebook's predict_and_evaluate function exactly.

Class labels: The checkpoint outputs 2 classes but does not store their names. They default to noise (index 0) and transits (index 1), matching torchvision's alphabetical ImageFolder ordering. If your training set used a different pair, update the CLASS_NAMES list at the top of frontend/model_service.py.

---

Project Structure

isro-bah-2026/
│
├── DESIGN.md                    # Visual & interaction design principles
├── LICENSE                      # GNU GPL v3
├── README.md
├── render.yaml                  # Render deployment blueprint
│
├── model/
│   ├── tessnet.ipynb            # Data preparation + ResNet18 training
│   └── sample_data/             # Example light-curve samples
│
├── paper/
│   ├── tessnet.tex             # LaTeX source of the project paper
│   └── tessnet.pdf             # Compiled paper
│
└── frontend/
    ├── app.py                   # Flask application & API routes
    ├── model_service.py         # Model loading, lazy init & inference
    ├── pipeline.py              # CSV/TIC → clean → BLS → fold → render → classify
    ├── gunicorn.conf.py         # Production WSGI server config
    ├── requirements.txt
    ├── tess_resnet18_model.pth  # Trained ResNet18 weights (~45 MB)
    │
    ├── assets/
    │   ├── icon.png             # App icon (favicon + loading splash)
    │   └── tessnet.pdf          # Paper, served via the "Read Paper" button
    │
    ├── templates/
    │   └── index.html           # Main dashboard template
    │
    └── static/
        ├── css/
        │   └── style.css        # Full design-system stylesheet
        └── js/
            ├── data.js          # Synthetic demo light-curve generator
            ├── charts.js        # Chart.js rendering + live-update API
            ├── help.js          # Help modal + fullscreen detail dialog
            ├── details.js       # Extended per-section reference content
            ├── predict.js       # Inputs → batch processing → live results + JSON export
            ├── onboarding.js    # First-visit welcome + "Analyze Your Data" CTA
            └── loader.js        # Splash / loading overlay

---

Installation

Requires Python 3.9+

# 1. Clone the repository
git clone https://github.com/<your-org>/isro-bah-2026.git
cd isro-bah-2026/frontend

# 2. Create and activate a virtual environment (recommended)
python -m venv env

# Windows
env\Scripts\activate
# macOS / Linux
source env/bin/activate

# 3. Install dependencies
pip install -r requirements.txt

requirements.txt already pins the CPU-only PyTorch wheels (via the PyTorch CPU index) to keep the install small. If you need a different build, install torch and torchvision manually, e.g.:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu

---

Usage

Development server

Start the Flask dev server from the frontend/ directory:

python app.py

Open http://127.0.0.1:5000 in your browser.

Production (gunicorn)

On Linux/Unix, serve the app with the bundled gunicorn config:

gunicorn -c gunicorn.conf.py app:app   # serves on 0.0.0.0:8000

The config uses 2 workers and a 120 s timeout (BLS period search and MAST downloads are slow), binds to $PORT, and forces matplotlib's headless Agg backend. Tune via WEB_CONCURRENCY, GUNICORN_TIMEOUT, and GUNICORN_BIND. (gunicorn does not run on native Windows — use the dev server or waitress there.)

Deploying to Render

The repo ships a render.yaml blueprint (root dir frontend/, start command gunicorn -c gunicorn.conf.py app:app). The index route is kept lightweight — the model and analysis libraries load lazily on first use, not on page load — so the worker boots fast. The full ML stack (torch + lightkurve + matplotlib) needs more RAM than the free tier, so the blueprint requests the Standard plan.

Running an analysis

The dashboard opens on a clearly-labelled sample view (synthetic data — nothing is stored between visits). In the Detection Result panel, pick one of three input modes:

Mode	Input	What happens
Image	A phase-folded light-curve PNG (black scatter, white background)	Classified directly by ResNet18; updates the verdict, confidence and probability bars.
CSV	A time,flux CSV	Server flattens → BLS period search → phase-folds → renders the model image → classifies, then repopulates every panel from the real data.
TIC ID	A TESS TIC ID (e.g. 25155310)	Downloads the SPOC light curve from NASA MAST and runs the full pipeline. Requires internet access.

Every mode accepts multiple inputs (several files, or a comma/space-separated list of TIC IDs): they process sequentially with a progress bar, the dashboard updates live per item, and all results export together as one JSON file.

REST API

Method	Endpoint	Description
GET	/	Serves the dashboard.
GET	/api/model-status	Returns model readiness and class labels as JSON.
POST	/predict	Multipart image upload → JSON classification result.
POST	/analyze/csv	Multipart file (time/flux CSV) → full pipeline result (params, classification, image, series).
POST	/analyze/tic	Form field tic_id → downloads from MAST and runs the full pipeline.

Example /predict response

{
  "ok": true,
  "predicted": "transits",
  "predicted_label": "Planetary Transit",
  "confidence": 0.97,
  "transit_detected": true,
  "probabilities": [
    { "name": "transits", "label": "Planetary Transit",  "p": 0.97 },
    { "name": "noise",    "label": "Noise / No Transit", "p": 0.03 }
  ]
}

---

License

Released under the GNU General Public License v3.0.