Hybrid physical–machine learning detection of polar vortex and sudden stratospheric warming

Supplementary Material – Δ±1 + Machine Learning Pipeline for Polar Vortex Event Detection

1. Overview

This supplementary document provides detailed methodological, computational, and data-handling information supporting the main pipeline described in the OSF Wiki post. It includes:

Synthetic dataset generation

Feature matrix construction

Machine learning pipeline configuration

Evaluation metrics and visualizations

Additional sensitivity and feature importance analysis

All results presented here are demonstration-phase outputs using synthetic data calibrated to the observational statistical structure. Real observational validation is forthcoming.

---

2. Synthetic Data Generation

2.1 Event Structure

Total rare events: n = 47 (~1.5% of dataset)

Event duration: 7–14 day pre-onset windows

Features generated:

Δ±1 index (physically motivated)

Proton flux

Electron flux

Ozone

Jet stream velocity

Random Gaussian noise to emulate observational variability

2.2 Matrix Construction

Rows: daily snapshots (1998–2024)

Columns: features + binary target (event/non-event)

Data split: 70% training / 30% test (temporal, no shuffle)

PCA-derived weights applied to Δ±1 index

---

3. Machine Learning Pipeline

3.1 Algorithm

XGBoost classifier

Gradient boosting to capture non-linear interactions

Hyperparameters (demonstration):

max_depth = 3

learning_rate = 0.1

n_estimators = 200

scale_pos_weight = adjusted for class imbalance

3.2 Validation

5-fold time-series cross-validation

Performance metrics: ROC-AUC, Precision-Recall AUC

Feature importance: mean gain across folds

---

4. Evaluation Results (Synthetic)

Metric Value

ROC-AUC 0.98
PR-AUC 0.92
Sensitivity Improvement (Δ±1 baseline) +18%
Feature Importance (mean gain) Δ±1: 42%
Proton flux: 25%
Electron flux: 18%
Ozone: 10%
Jet velocity: 5%

Note: These results are synthetic demonstration outputs; real observational AUC expected to be lower due to weaker signal and noise.

---

5. Sensitivity Analysis

Local perturbation tests: each feature incremented ±1σ

Proton flux had highest impact on predicted probability

Δ±1 index remained dominant across perturbation tests

Minor contributors: ozone, jet velocity

---

6. Methodological Validation Checks

Correct feature alignment with events confirmed

Temporal data split validated (no leakage)

Synthetic signal strength verified to reproduce physical hierarchy

Pipeline reproducible: all scripts provided

---

7. Next Steps for Observational Application

8. Build hindcast dataset (1998–2024) with real satellite measurements (NOAA-POES, MetOp, GOES, CSES).

9. Implement rolling-window validation for realistic temporal prediction.

10. Add baseline models: logistic regression, climatology.

11. Evaluate calibration: Brier score, calibration plots.

12. Prospective validation during Solar Cycle 25 (2025–2035).

---

8. Data & Code Availability

Synthetic datasets and scripts: OSF DOI 10.17605/OSF.IO/WXBDA

Full reproducible pipeline for methodological demonstration included

Observational data integration planned for follow-up

---

9. References

10. Pérez Pulido, C.J., Integrating Space Weather and Atmospheric Physics: Δ±1 + ML Pipeline for Polar Vortex Detection, OSF, 2026, DOI: 10.17605/OSF.IO/WXBDA

11. NOAA-POES, MetOp, GOES, CSES – public satellite datasets (1998–2024)

12. XGBoost Documentation: Chen, T. & Guestrin, C., XGBoost: A Scalable Tree Boosting System, KDD 2016

Perfecto, aquí tienes una versión

Integrating Space Weather and Atmospheric Physics: Δ±1 + Machine Learning for Polar Vortex Event Detection

Overview

Extreme stratospheric events, such as sudden stratospheric warmings (SSWs) and polar vortex disruptions, have significant impacts on mid-latitude weather, energy systems, and cold-weather risk. Predicting these rare events is difficult because of complex interactions between atmospheric dynamics, chemistry, and energetic particle flux from space weather. Traditional operational models (e.g., ECMWF, NOAA) capture large-scale circulation but often miss non-linear couplings, particularly those involving energetic particle precipitation and stratospheric chemistry.

This project presents a hybrid approach combining:

1. Δ±1 Physical Index – a physically motivated planetary energy-state measure summarizing proton flux, electron flux, thermal gradients, and atmospheric chemistry.

2. Gradient-boosted Machine Learning (XGBoost) – captures residual non-linear interactions not explicitly represented in physics-based models.

Objectives

Develop a reproducible hybrid pipeline integrating physics-based indices with machine learning for rare stratospheric event detection.

Validate methodological feasibility using synthetic datasets calibrated to observational statistics.

Provide a structured framework for future application to hindcast (1998–2024) and real-time observational datasets.

Encourage community replication and methodological critique via publicly available scripts and synthetic data.

Methods

Δ±1 Physical Index

The Δ±1 index is formulated as:

ΔE_system = Σᵢ wᵢ * Ep,i + Σⱼ vⱼ * Ee,j − k * ∇T

Proton flux (Ep), electron flux (Ee), and thermal gradients (∇T) are combined with PCA-derived weights (wᵢ, vⱼ) to form a bounded energy-state variable.

Atmospheric chemistry (O₃, NOx, etc.) and jet stream velocity are included as modifiers.

The index provides interpretable, physically grounded insight into energy coupling between magnetosphere, ionosphere, and stratosphere.

Machine Learning Integration

Algorithm: XGBoost with gradient boosting for rare event classification.

Input features: Δ±1 index, proton flux, electron flux, ozone, jet stream velocity, and additional atmospheric parameters.

Synthetic demonstration:

n = 47 rare events (~1.5% frequency)

Temporal split (train/test) without shuffling

Evaluation: ROC-AUC ≈ 0.98 (synthetic data; real observational AUC expected to be lower)

Feature importance pattern (mean gain, synthetic):

Δ±1 index: 40–45%

Proton flux: 25–30%

Electron flux: 15–20%

Ozone: 10%

Jet stream velocity: 5%

Validation Strategy

1. Synthetic data validation ensures pipeline architecture works before observational application.

2. Next step: hindcast observational dataset (1998–2024) using rolling-window evaluation for realistic predictive performance.

3. Metrics: ROC-AUC, PR-AUC, Brier score, calibration plots.

4. Prospective real-time validation during Solar Cycle 25 (2025–2035) as the decisive test.

Significance

Provides a complementary predictive layer to existing operational models.

Sensitive to energetic particle–chemistry coupling that is typically underrepresented.

Offers potential early warning (7–14 days) for extreme stratospheric events, aiding energy planning, seasonal forecasting, and cold-weather risk assessment.

Data and Code Availability

All synthetic datasets and computational scripts are publicly available for replication and methodological critique.

Future observational hindcast datasets will be integrated upon validation.

Open Questions for the Community

Which observational proxies best capture energetic particle precipitation effects on stratospheric ozone?

How should class imbalance for rare stratospheric events be optimally handled in ML frameworks?

Are there existing hindcast datasets that could accelerate observational validation?

References & Resources

Main paper (demonstration phase): DOI 10.17605/OSF.IO/WXBDA

Supplementary material: DOI 10.17605/OSF.IO/WXBDA

ISHEA Institute program overview: DOI 10.17605/OSF.IO/JBNGT

Related work: Integrating Planetary Energy Flows and Solar Cycles DOI 10.17605/OSF.IO/GY83R