🪐 JyotirVega Exoplanet Detection Pipeline

ISRO Bharatiya Antariksh Hackathon 2026 — Problem Statement 7

Team JyotirVega | Aurixys | LOGMIEER, Nashik | SPPU

---

What this is

A complete, 10-stage AI pipeline that downloads real TESS light curves from NASA MAST, screens for stellar variability, detrends with Wotan biweight, searches for transits with Transit Least Squares, vets false positives with 5 independent physics tests, classifies signals with a dual-view CNN, and fits transit parameters with batman + emcee MCMC.

Every PS-7 objective is addressed:

PS-7 Requirement	Implementation
Identify periodic dips	TLS limb-darkened transit search
Classify into transit/eclipse/blend/other	Dual-view CNN + rule ensemble + variability screen
Apply to science datasets	--sector batch mode, --tics mode
SNR / significance	Folded SNR + TLS SDE + bootstrap FAP
Period, depth, duration estimates	batman MAP → emcee MCMC posteriors
Confidence level	Temperature-scaled calibrated probability
3-page report	Auto-generated HTML + Markdown
Visualization	6-panel diagnostic + corner plots + population

---

Windows Quick Start

1. Double-click setup_windows.bat    ← installs everything
2. Double-click run_demo.bat         ← runs pipeline, no internet needed
3. Double-click run_dashboard.bat    ← opens web dashboard at localhost:8501

---

Linux / Mac Quick Start

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

# Demo (no internet)
python run_pipeline.py --demo

# Dashboard
python -m streamlit run dashboard/app.py

# Train CNN
python train_model.py --synthetic --epochs 50 --n-per-class 300 --eval

---

All commands

# Synthetic demo (no internet, fast)
python run_pipeline.py --demo

# Demo with MCMC (slower, full posterior uncertainties)
python run_pipeline.py --demo --mcmc

# Real TESS targets (needs internet)
python run_pipeline.py --tics 261136679 388857263 100100827

# Full sector batch (20-30k targets)
python run_pipeline.py --sector 1 --max-targets 100

# Known targets demo
python run_pipeline.py --real-demo

# Train CNN — synthetic (fast, ~5 min)
python train_model.py --synthetic --epochs 50 --n-per-class 300 --eval --calibrate

# Train CNN — real TESS data (best accuracy, takes hours)
python train_model.py --real --epochs 100 --n-per-class 500 --eval --calibrate

# Build independent validation set
python scripts/build_validation_set.py --n-per-class 50

# Evaluate against ground truth
python scripts/evaluate_validation.py \
    --predictions outputs/results_table.csv \
    --ground-truth data/validation/validation_targets.csv

# Interactive dashboard
python -m streamlit run dashboard/app.py

---

Pipeline stages

Stage	Module	What it does
1	data_loader.py	Download real TESS SPOC 2-min LC from MAST
2	stellar_variability.py	Lomb-Scargle starspot pre-screen
3	preprocessor.py	Wotan biweight flatten + sigma clip
4	detector.py + multiplanet.py	TLS search + iterative multi-planet
5	snr_calculator.py	Folded SNR + TLS SDE
6	significance.py	Bootstrap permutation FAP
7	false_positive.py	5 FP tests (odd/even, secondary, centroid, V-shape, depth-var)
8	blend_crosscheck.py	TIC contamination + Gaia DR3 neighbours
9	classifier.py	Dual-view CNN (2001-pt + 201-pt) + rule ensemble
10	fitter.py	batman MAP → emcee MCMC → posterior uncertainties

---

Output files (all in outputs/)

File	Content
TIC_*_diagnostic.png	6-panel per-star plot
TIC_*_corner.png	MCMC posterior corner plot
population_summary.png	Period-depth scatter + SNR histogram
results_summary.json	Machine-readable results with all parameters
results_table.csv	Spreadsheet-friendly summary
report.html	3-page ISRO submission report
report.md	Markdown version
demo_results.csv	Demo run accuracy table
pipeline.log	Full execution log

After training:

File	Content
models/classifier.h5	Trained CNN weights
models/calibration_temperature.json	Temperature scaling + reliability table
outputs/confusion_matrix.png	Classification accuracy heatmap
outputs/training_history.png	Loss + accuracy curves

---

Known test TIC IDs

TIC ID	Object	Expected class
261136679	HD 21749b	planet_transit
388857263	Known EB	eclipsing_binary
100100827	TOI candidate	planet_transit

---

Architecture (what makes this different)

1. TLS over BLS — physically accurate limb-darkened template, ~10-15% better sensitivity
2. Wotan over SG — robust biweight detrending, no polynomial artifacts
3. Starspot screen — explicit Lomb-Scargle variability routing before CNN
4. Dual-view CNN — global (2001-pt) + local (201-pt) views, mirrors AstroNet
5. Bootstrap FAP — assumption-free permutation significance, not analytic approximation
6. Blend cross-check — TIC contamination + Gaia DR3 neighbour, not centroid alone
7. MCMC posteriors — full credible intervals, not point estimates
8. Calibrated confidence — temperature scaling, reliability table, not raw softmax
9. Multi-planet search — iterative mask + re-search
10. Independent validation — accuracy from real catalogs not training distribution

---

References

• Shallue & Vanderburg (2018) — AstroNet dual-view CNN
• Hippke & Heller (2019) — Transit Least Squares
• Hippke et al. (2019) — Wotan detrending
• Kreidberg (2015) — batman transit model
• Foreman-Mackey et al. (2013) — emcee MCMC
• Prša et al. (2022) — TESS EB catalog
• Guo et al. (2017) — neural network confidence calibration

---

Team JyotirVega | Aurixys | LOGMIEER, Nashik | SPPU PI: Vishal Shivaji Patil