Speaker
Description
Spacecraft do not sample Earth’s magnetosphere uniformly. Measurements cluster along orbital tracks and mission-targeted regions, making it hard to build global maps of plasma environments directly from data. In this talk, I will describe a fully data-driven approach for constructing a magnetospheric atlas from a decade of NASA Magnetospheric Multiscale (MMS) observations. We first compress the continuous time series into short states (2-minute windows) that capture both typical conditions and variability using summary statistics of key plasma moments and magnetic-field features. To counter spatial sampling imbalance, we discretize space with an adaptive 3D octree that refines only where observations are dense, and we use octree-informed weighting to reduce orbit-driven bias during learning.
On top of this representation, we train a discrete generative model (a vector-quantized variational autoencoder) to discover a small set of recurring plasma regimes without expert thresholds or hand labels. The resulting atlas maps where each regime occurs and, crucially, provides per-regime feature distributions. I will show how the learned regimes align with familiar structures (e.g., sheath, boundary layers, plasma sheet, lobes) and how their occupancy shifts between quiet and storm-time conditions, illustrating the map as both an interpretive tool and a foundation for automated labeling, anomaly detection, and generative expected conditions baselines
| Numerical model | Generative modeling |
|---|
