🪐 Exo Gargantua: TESS Exoplanet ML Vetting & Detection Pipeline

---

📖 Table of Contents

• Executive Overview
• Key Features & Major Upgrades (v2.0)
• Astrophysical & ML Architecture
• Vetting Engine & Difference Image Analysis (DIA)
• Machine Learning Vetting Architecture
• Synthetic Injection-Recovery Framework
• Benchmark Results: TIC 25155310 (WASP-126b)
• Installation & Environment Setup
• Command-Line Interface (CLI) Usage
• Testing & Continuous Integration (CI)
• Authors & Citation

---

🔭 Executive Overview

Exo Gargantua is a publishable, research-grade pipeline engineered to process millions of TESS cadences. It transitions exoplanet discovery from brittle heuristics into a unified statistical and Machine Learning engine.

Starting from raw SPOC (Science Processing Operations Center) calibrated Target Pixel Files (TPFs) and Light Curves downloaded programmatically from the Mikulski Archive for Space Telescopes (MAST), the pipeline performs:

1. Multi-Sector Ingestion & Stitching with crowding metric (CROWDSAP) corrections.
2. Continuum-Referenced Transit-Masked Detrending to preserve true transit depths.
3. Decimated Box Least Squares (BLS) periodic transit searches with bootstrap False Alarm Probability (FAP) estimation.
4. Dynamic Stellar Limb-Darkening Estimation via ldtk and the TESS Input Catalog (TIC v8.2).
5. Pixel-Level Difference Image Analysis (DIA) to detect background blended contaminants down to sub-pixel accuracy.
6. Supervised ML Vetting via a calibrated Random Forest classifier trained on spatial and photometric false-positive models.
7. Bayesian MCMC Posterior Sampling using emcee and batman with full Monte Carlo physical parameter propagation ($R_p$, $a$, $T_{\text{eq}}$, $S_{\text{eff}}$).

---

🚀 Key Features & Major Upgrades (v2.0)

Capability	Pipeline v1 (Baseline)	Pipeline v2 (Current)
Vetting Paradigm	Hard heuristic thresholds	Supervised Random Forest Classifier + Physical DIA Vetoes
Centroid Analysis	Simple flux-weighted 1st moments	True Difference Image Analysis (DIA) Out-of-Transit vs In-Transit
Limb Darkening	Static quadratic constants ($u_1=0.3, u_2=0.1$)	Dynamic $q_1, q_2$ profiles computed via ldtk from TIC ($T_{\text{eff}}, \log g, [\text{Fe}/\text{H}]$)
Validation Suite	Manual targets / Error-prone harness	Full Synthetic Transit Injection-Recovery & Multi-Modal Dataset Generator
Packaging & CI	Script collection (requirements.txt)	Modern pyproject.toml PEP 518/621 standard + GitHub Actions CI
Reporting & Telemetry	Basic terminal printouts	Structured JSON Reports, Corner Plots, and 4-Panel Publication Figures

---

🏗️ Astrophysical & ML Architecture

flowchart TD
    A[MAST API / SPOC Lightcurves & TPFs] --> B[Phase 1: Ingestion, Stitching & Crowding Correction]
    B --> C[Phase 2: Continuum-Referenced Transit-Masked Detrending]
    C --> D[Phase 3: Fast Decimated BLS Search & Bootstrap FAP]
    
    D --> E[Phase 3b: Physical Vetting Engine]
    E --> F[Odd/Even Transit Depth Delta Check]
    E --> G[Rolling-Mean Red Noise Secondary Eclipse Test]
    E --> H[Difference Image Analysis DIA Centroid Shift]
    
    F --> I[Feature Extraction Vector]
    G --> I
    H --> I
    D --> I
    
    I --> J[Phase 3c: Calibrated ML Vetter Random Forest]
    J -->|Disposition & Planet Probability| K{Pass Physical & ML Vetting?}
    
    K -->|Yes / Validated Candidate| L[Phase 4: Dynamic Limb Darkening LDTk + MCMC emcee]
    K -->|No / Blended False Positive| L
    
    L --> M[Phase 4b: Monte Carlo Physical Parameter Derivation]
    M --> N[Phase 4c: Publication Figures & JSON Summary Report]

---

🔬 Vetting Engine & Difference Image Analysis (DIA)

The vetting suite isolates genuine planetary transits from astrophysical false positives (Eclipsing Binaries, Grazing Binaries, and Background Eclipsing Binaries):

1. Difference Image Analysis (DIA) Centroid Vetting

To verify that the transit signal originates from the target star rather than a nearby background star within the large TESS pixel aperture ($\approx 21''/\text{pixel}$):

• Construct the reference out-of-transit image $\bar{I}{\text{out}}$ and in-transit image $\bar{I}{\text{in}}$.
• Compute the differential flux image: $\Delta I = \bar{I}{\text{out}} - \bar{I}{\text{in}}$.
• Calculate the flux-weighted centroid of the differential dip $\vec{r}{\text{res}} = (\bar{x}{\text{res}}, \bar{y}{\text{res}})$ and the reference star $\vec{r}{\text{ref}} = (\bar{x}{\text{ref}}, \bar{y}{\text{ref}})$.
• Calculate the Euclidean centroid shift: $$\Delta r = \sqrt{(\bar{x}{\text{res}} - \bar{x}{\text{ref}})^2 + (\bar{y}{\text{res}} - \bar{y}{\text{ref}})^2}$$
• Veto Threshold: A shift $\Delta r \ge 0.333,\text{pixels}$ ($\approx 7''$) triggers an immediate FAILED_CENTROID_DIA flag.

2. Odd/Even Depth Consistency

Calculates the absolute depth delta $\Delta \delta = |\delta_{\text{odd}} - \delta_{\text{even}}|$. Deep deltas ($\Delta \delta > 0.005$) indicate alternating primary and secondary eclipses characteristic of eclipsing binary systems.

3. Integrated Red-Noise Secondary Eclipse Check

Computes the rolling-mean standard deviation in out-of-transit chunks matched to the transit duration, scoring secondary eclipse significance: $$\sigma_{\text{sec}} = \frac{\delta_{\text{sec}}}{\sigma_{\text{rolling_means}}}$$ Any secondary eclipse with $|\sigma_{\text{sec}}| > 3.0\sigma$ flags the candidate as an occulting stellar companion.

---

🤖 Machine Learning Vetting Architecture

Rather than relying purely on heuristic threshold cuts, exoplanet_pipeline/ml_vetting.py implements a calibrated Random Forest Classifier trained on multi-modal feature vectors:

$$\vec{x} = \left[ \text{SNR}{\text{BLS}}, \Delta \delta{\text{odd-even}}, \sigma_{\text{sec}}, \Delta r_{\text{DIA}}, \text{TSNR}, \mathbb{I}(\Delta r \ge 0.333) \right]$$

Feature Importances in Trained Model:
  centroid_shift           : 29.65%
  depth_diff               : 27.82%
  centroid_fail (DIA Veto) : 22.54%
  secondary_eclipse_sigma  : 17.76%
  snr                      :  1.31%
  bls_power                :  0.92%

Spatial Veto & Consistency Guarantee

Because 1D transit injection cannot synthesize 2D pixel-level centroid shifts into Target Pixel Files, the training pipeline couples synthetic transit datasets with multi-modal physical false-positive generators (Background Eclipsing Binaries with centroid offsets $\mathcal{U}(0.35, 3.0),$px, and EBs with high secondary eclipses).

Furthermore, MLVetter.predict() enforces hard physical gating: a candidate failing the spatial DIA test ($\Delta r \ge 0.333,$px) has its planet probability suppressed to $<0.5%$, preventing high 1D SNR from overriding spatial contamination.

---

🧪 Synthetic Injection-Recovery Framework

The pipeline includes exoplanet_pipeline/validate_injection.py to evaluate pipeline recovery completeness across the parameter space ($P \in [1, 15],$d, $R_p/R_\star \in [0.01, 0.15]$):

• Ingests real quiet TESS light curves.
• Injects non-linear Mandel & Agol (2002) transit models using batman.
• Runs blind detrending and BLS detection to evaluate recovery fraction and signal preservation.

---

📊 Benchmark Results: TIC 25155310 (WASP-126b)

The complete end-to-end pipeline was executed on benchmark target TIC 25155310 (WASP-126b).

Extracted Parameters (output/TIC_25155310_summary.json)

Parameter	Pipeline Value	Uncertainty ($-1\sigma / +1\sigma$)	Physical Unit
Orbital Period ($P$)	3.288790	$\pm 0.000001$	days
Transit Epoch ($t_0$)	1327.519196	$-0.000448 / +0.000387$	BTJD
Radius Ratio ($R_p / R_\star$)	0.063223	$-0.000394 / +0.000926$	—
Semi-Major Axis ($a / R_\star$)	8.007972	$-0.511112 / +0.147525$	—
Orbital Inclination ($i$)	88.52	$-1.65 / +0.99$	degrees
Planetary Radius ($R_p$)	8.77	$\pm 0.44$	$R_\oplus$ ($0.78,R_{\text{Jup}}$)
Semi-Major Axis ($a$)	0.0449	$\pm 0.0007$	AU
Equilibrium Temperature ($T_{\text{eq}}$)	1359	$\pm 42$	K
Centroid Shift (DIA)	0.4965	—	pixels (passed: false)
ML Vetting Disposition	FALSE_POSITIVE	prob: 0.00%	Flagged: FAILED_CENTROID_DIA

---

💻 Installation & Environment Setup

Prerequisites

• Python $\ge 3.9, \le 3.11$
• Git

Installation via pyproject.toml

# 1. Clone the repository
git clone https://github.com/bytes06runner/exo_gargantua.git
cd exo_gargantua

# 2. Create and activate a virtual environment
python3 -m venv exo-env
source exo-env/bin/activate

# 3. Install in editable mode with development & testing dependencies
pip install -e ".[dev]"

---

🛠️ Command-Line Interface (CLI) Usage

1. Run the Full End-to-End Pipeline

Process a TESS target through ingestion, detrending, BLS search, vetting, DIA, and MCMC parameter estimation:

python run_pipeline.py "TIC 25155310" --report

Options:

• --report: Generates summary JSON in output/ and publication 4-panel diagnostic plots.
• --no-mcmc: Skips MCMC posterior estimation for fast screening.
• --no-centroid: Skips Target Pixel File (TPF) downloads.

2. Generate Synthetic Injection-Recovery Datasets

python exoplanet_pipeline/validate_injection.py

3. Train the ML Vetting Model

python exoplanet_pipeline/ml_vetting.py

4. Run Astrophysical Ground-Truth Validation Harness

python run_pipeline.py --validate

---

🧪 Testing & Continuous Integration (CI)

The repository includes a comprehensive pytest test suite configured with GitHub Actions CI on Ubuntu runners with Python 3.10 and 3.11:

# Execute local unit tests
pytest tests/ -v

Test Coverage Includes:

• Kepler's Third Law astronomical scaling and error boundaries (tests/test_math.py).
• Rolling-mean red-noise boxcar standard deviations (tests/test_math.py).
• Odd/Even transit depth deltas (tests/test_math.py).
• ML Vetter model training, prediction, and spatial DIA centroid veto consistency (tests/test_ml_vetting.py).

---

👤 Authors & Citation

Srijeet Prasad Banerjee (@bytes06runner)
Contact: cybrobasics@gmail.com

License

This project is open-source and licensed under the MIT License.