EXO-CHECKMATE ♟️

A Physics-Informed "Grandmaster" Engine for Exoplanet Discovery

Watch Demo Video • Read The Paper (Coming Soon)

---

♟️ The Concept

Space exploration is a game of probability. Current AI models are "Time-Blind"—they treat light curves as static images, often confusing Eclipsing Binaries (EBs) with real planets because they share similar shapes in low resolution.

EXO-CHECKMATE treats discovery like a game of Chess. We replaced brute force with strategy:

1. THE STRATEGIST (Layer 0): Immediate pruning of stars that are too large, too hot, or too noisy.
2. THE TACTICIAN (Layer 1): Finds the rhythm (periodicity) using Box Least Squares (BLS).
3. THE VISIONARY (Layer 2): A Hypersonic Dual-View Engine that sees both Shape and Time.
4. THE JUDGE (Layer 3): Final validation using Hill Stability Mechanics (Gladman, 1993).

---

🧬 Architecture

We implement a physics-informed pipeline that funnels terabytes of TESS data into a "Checkmate" verdict.

graph TD
    %% NODES
    Data[(TESS Data)] --> L0{Layer 0: Strategist}
    
    subgraph Pipeline [The Grandmaster Pipeline]
        L0 -->|Reject > 1.5 R_sun| Trash[Bin]
        L0 -->|Accept| L1[Layer 1: Tactician]
        
        L1 -->|Scientific Binning + BLS| Signal{Signal Found?}
        Signal -->|No| Trash
        Signal -->|Yes| L2[Layer 2: Visionary]
        
        subgraph ML [The Hypersonic Engine]
            L2 --> CNN[Head A: CNN\nShape Expert]
            L2 --> LSTM[Head B: LSTM\nSequence Expert]
            L2 --> PHYS[Head C: Dense\nPhysics Prior]
            CNN & LSTM & PHYS --> FUSION((FUSION))
        end
        
        FUSION -->|Score > 95%| L3[Layer 3: The Judge]
        
        L3 -->|Hill Stability Check| Stable{Orbit Stable?}
        Stable -->|No| Trash
        Stable -->|Yes| FINAL[✅ CHECKMATE\nConfirmed Candidate]
    end

Loading

---

🧠 The "Hypersonic" Engine

Standard CNNs fail on TESS data because they lack temporal context. Our Dual-View Hybrid Neuro-Probabilistic Engine solves this:

• Head A (Shape Expert): A 1D-CNN optimized for Spatial Features. It distinguishes the "U-Shape" of a planetary transit from the "V-Shape" of a binary star.
• Head B (Sequence Expert): A Bi-Directional LSTM that analyzes the Global View. It detects Transit Timing Variations (TTVs) and secondary eclipses that rule out planetary candidates.
• Head C (Physics Prior): A dense layer that injects raw orbital mechanics (Period, Density, SNR) directly into the decision latent space.

Dataset: Trained on Kepler Labelled Time Series + 10,000 Synthetic Injections (The "Frankenstein" Method) to handle class imbalance.

v2.3 Upgrades (Jan 2026):

• Translational Invariance: Random jitter added to training crops to prevent center-pixel overfitting.
• Scientific Validity: Quadratic Limb Darkening (ExoCTK physics) replaces simple box models.
• Physics Normalization: Deep Learning inputs scaled to [0,1] range to ensure fair weighting of weak signals (e.g. Period) vs strong signals (e.g. SNR).
• Efficiency: Strided Convolution LSTM inputs reduce sequence length 1000->250 for sharper gradient flow.

---

✅ Model Verification

To verify the integrity of the Hypersonic Engine (v2.3) on your machine, run the included verification script:

cd ml_engine
python verify_model_v3.py

This tests the model against:

1. Simulated Hot Jupiter (Should pass with >99% confidence)
2. Eclipsing Binary (V-Shape) (Should be REJECTED <10% confidence)
3. Pure Noise (Note: Model requires BLS pre-filtering; raw flat noise may produce artifacts, but Layer 1 filters this out).

---

🔭 The "Glass Box" Dashboard

We don't ask you to trust the AI. We show you the math.

(Note: Screenshot represents the tactical view showing Folded Curves, Periodograms, and the Hypersonic Decision Snapshot)

• Scientific Binning: Downsamples 20,000+ data points to 1,500 for 60fps rendering without loss of signal fidelity.
• Hypersonic Snapshot: Relays real-time confidence scores split by "Shape" (CNN) vs "Temporal" (LSTM) components.
• Goldilocks Radar: Real-time calculation of the Habitable Zone based on equilibrium temperature and stellar radius.

---

⚡ Installation

Prerequisites

• Python 3.10+
• Node.js 18+

1. Clone & Install Backend

git clone https://github.com/your-team/exo-checkmate.git
cd exo-checkmate/backend
pip install -r requirements.txt
python main.py

2. Install Frontend

cd ../frontend/cyberpunk-nasa
npm install
npm run dev

Visit http://localhost:3000 to access Mission Control.

---

📜 Citations & Research

Our work builds upon the following foundational papers:

1. ExoMiner++ 2.0: Deep Learning for Exoplanet Validation (arXiv:2601.14877)
2. Mandel & Agol (2002): Analytic Light Curves for Planetary Transit Searches
3. Gladman (1993): Dynamics of Systems of Two Close Planets (Hill Stability Criterion)
4. TESS Science Pipeline: NASA High Energy Astrophysics Science Archive Research Center

---

Built for DuHacks 5.0 🚀 Because finding Earth 2.0 shouldn't be boring.