This project addresses the detection of fast-ion plasma flow events in the Earth's nightside magnetosphere using high-frequency satellite magnetic field data. The focus is on building a volatility-informed event detection pipeline that leverages both domain-specific statistical features and deep learning techniques.

Fast-ion flow events are critical to understanding magnetospheric dynamics and substorm activity. However, due to their rarity and the high-frequency nature of the underlying satellite data, identifying these events in real time remains a challenge.

This project proposes a scalable, data-driven approach for detecting fast-ion flow intervals by analyzing volatility patterns in magnetic field fluctuations. The resulting model demonstrates strong event capture rates and generalization across different satellites.

• Mission: NASA THEMIS (Time History of Events and Macroscale Interactions during Substorms)
• Satellites Used: Themis A, D, and E
• Sampling Rate: 3-second frequency
• Data Volume: ~60 million rows of magnetic field measurements
• Features Used:
  • $B_x, B_y, B_z$ magnetic field components
  • Derived volatility features:
    • AR(1)-GARCH(1,1) volatility estimates per component
    • 9-window rolling standard deviations per component

All features were scaled to [0, 1] using min-max normalization per satellite. Labels for fast-ion flow events were constructed based on velocity thresholds, applied independently per satellite.

• Link for cleaned data: Box Link

Traditional event detection models often rely on hand-tuned thresholds or statistical baselines that fail under dynamic plasma conditions. Our goal was to build a model that could learn temporal volatility patterns associated with fast-ion flows—without directly using velocity as an input.

Volatility was chosen as the modeling basis because:

• It reflects local plasma instability and turbulence
• It is highly correlated movement with ion velocity
• It provides richer temporal dynamics than raw magnitude

A hybrid approach was used that combines physics-aware statistical feature engineering with deep sequence modeling:

• Feature Engineering:

  • Volatility modeled using AR(1)-GARCH(1,1) to capture memory effects. The AR(1)-GARCH(1,1) can be interpreted as a stochastic difference equation – a discrete-time analog to stochastic differential equations. This allows us to estimate both the drift and diffusion behavior of the magnetic field time series.
  • Localized fluctuation captured via 9-timestep rolling standard deviations

• Neural Network Architecture: A custom deep learning model, TimeSeqNet, was designed to detect sequences of volatility patterns preceding fast-ion events.

  • Convolutional layer for local pattern extraction
  • Bidirectional LSTM layer for temporal context
  • Multi-head attention for per-timestep interpretability
  • Sequence-to-sequence output structure with aggregation at the interval level
  • Multi-Output Shared Representation Learning (MOSRL) to capture per timestep information and latent feature representations for event sequences
  • Trained using a custom loss function with Focal Binary Crossentropy and Tversky loss to address class imbalance for per-timestep predictions, and uses Huber loss for the event ratio output.

The model takes in sequences of engineered features and outputs per-timestep probabilities, which are aggregated to produce a final event classification for each interval.

• Training/Test Splits:

  • Satellite-randomized and time-respecting splits were used
  • Additional generalization testing: train on A+D, test on E

• Metrics:

  • Precision, recall, and F1 score computed at multiple thresholds
  • Event-level evaluation using captured vs. missed intervals
  • Missed to Captured Events ratio over time
  • Data drift analysis

• Best Performance (10 Epochs, F1-Optimal Threshold = 0.58):

  • Precision: 0.76
  • Recall: 0.83
  • F1 Score: 0.79

The model was found to generalize well across satellites and years, with consistent recall and stability over time.

This project demonstrates that volatility-driven modeling is an effective strategy for detecting fast-ion plasma flow events in the magnetosphere. By leveraging GARCH and rolling standard deviation features alongside a custom sequence learning architecture, we achieve robust, generalizable event detection on high-frequency satellite data.