Aircraft Wing Corrosion Risk Prediction

HAKS 2026 Challenge — Predict corrosion risk for aircraft wings using environmental data.

🎯 Quick Start

# 1. Setup environment (creates venv + installs dependencies)
./scripts/setup.sh

# 2. Activate virtual environment
source venv/bin/activate

# 3. Run cross-validation
python src/corrosion_model.py

# 4. Generate submission file
python src/corrosion_model.py --submit

📊 Current Performance

Model	Brier Score	Improvement vs Baseline
Ensemble	0.124 ± 0.010	50.4%
HistGradientBoosting	0.129 ± 0.015	48.4%
Logistic Regression	0.170 ± 0.011	32.0%

Baseline: 0.25 (constant 0.5 prediction)

📁 Project Structure

.
├── input/                          # Data files (not in git)
│   ├── environment_training.csv    # 63.5k rows, 758 aircraft
│   ├── corrosions_training.csv     # 790 corrosion events
│   ├── environment_test.csv        # 142 test aircraft
│   └── sample_submission.csv       # 164 predictions needed
├── output/                         # Generated submissions (timestamped)
├── src/
│   └── corrosion_model.py         # Main model (THE scored file)
├── scripts/
│   ├── setup.sh                   # Environment setup
│   └── run.sh                     # Run helper (cv|submit)
└── Docs/
    └── ANALYSIS_AND_IMPROVEMENTS.md  # Detailed technical analysis

🔬 How It Works

Problem

Predict corrosion_risk ∈ [0,1] per (aircraft, reference_month) using:

• Environmental data: weather (METAR), aerosols (Copernicus), parking time
• Label construction: corrosion month → 1, exactly 24 months before → 0
• Key insight: Negatives share same calendar month (seasonality neutralized)

Pipeline

Environment Data → Feature Engineering → Model Training → Prediction
                   (age + cumulative     (Ensemble:
                    exposure features)    60% GBDT + 40% Logistic)

Features (77 total)

• Age: age_months since first observation
• Cumulative exposure: cummean_* (expanding mean of 33 environmental drivers)
• Integrated dose: cumsum_dose_* (parking time × corrosive factors)
• Current values: cur_* (raw monthly values)

Validation

• GroupKFold by aircraft_id (5 splits)
• Prevents temporal leakage (aircraft never split across train/val)
• Metric: Brier score (lower is better)

🚀 Advanced Usage

Run specific modes

./scripts/run.sh cv       # Cross-validation only
./scripts/run.sh submit   # Generate submission file

Direct Python execution

python src/corrosion_model.py            # CV evaluation
python src/corrosion_model.py --submit   # Generate submission

📈 Improvement Roadmap

See Docs/ANALYSIS_AND_IMPROVEMENTS.md for detailed analysis. Key opportunities:

1. Recency features (+10-15% expected improvement)

   • Rolling window doses (12/24 months)
   • Exponentially decay-weighted exposure

2. Age extrapolation fix (prevents wild predictions on old test aircraft)

   • Cap age_months at training maximum

3. Hyperparameter tuning (+2-5% expected)

   • Systematic search with GroupKFold CV

4. Optimized ensemble weight (+1-2% expected)

   • Currently fixed at 0.6, should be tuned

🛠️ Technical Details

Dependencies

• Python 3.8+
• pandas, numpy, scikit-learn
• See requirements.txt for versions

Data Format

• Training: 758 aircraft, monthly environmental history (2014-04 to 2026-05)
• Test: 142 aircraft (delivered 2014, older than training)
• Labels: 1270 simple labels (616 positive, 654 negative)

Model Architecture

• HistGradientBoostingClassifier: Handles missing values natively, fast
• Logistic Regression: Linear baseline with StandardScaler
• Ensemble: Weighted average (60/40 split)
• Calibration: Isotonic regression on out-of-fold predictions
• Post-processing: Structural prior (later month → higher risk)

📝 Notes

• No temporal leakage: All features computed up to (inclusive) reference month
• Trust CV, not public leaderboard: GroupKFold gives honest performance estimate
• Test set challenge: Older aircraft than training (extrapolation risk)
• Outputs: Timestamped in output/submission_YYYYMMDD_HHMMSS.csv

🔗 References

• Challenge: HAKS 2026 (IBM × Airbus × AWS)
• Metric: Brier score
• Baseline: 0.25 (constant prediction)
• Current best: 0.124 (50% improvement)

---

Last updated: 2026-06-11