Finding Novel Precursors for Solar Wind Stream Interaction Regions With Interpretable Deep Learning
4 affiliations
- 1Space Weather TREC University of Colorado Boulder CO USA
- 2Department of Physics and Astronomy University of Iowa Iowa IA USA
- 3Physical Research Laboratory Ahmedabad India
- 4Heliophysics Science Division NASA Goddard Space Flight Center Greenbelt MD USA
Abstract
Solar wind stream interaction regions (SIRs) drive recurrent geomagnetic storms, yet most existing catalogs rely on expert inspection and simple thresholds that are subjective and can miss events with complex morphologies. We present SIREN (SIR Encoder Network), a lightweight Transformer‐based model for per‐timestep SIR detection from in situ solar wind observations. The model ingests sequences of 11 solar wind parameters, spanning magnetic field, velocity, and thermodynamic properties. With approximately 100,000 trainable parameters in a two‐layer encoder architecture, SIREN is trained using weighted binary cross‐entropy loss and a cosine annealing learning rate. Platt scaling is applied to produce well‐calibrated detection probabilities. On a held‐out test set of 96 events, the calibrated model achieves a high skill score and analysis of the self‐attention weights confirms that the model concentrates on the SIR, grounding its decisions in the physically relevant portion of each sequence. Integrated Gradients attribution reveals a quantifiable feature hierarchy: proton density (23.4%) and magnetic field magnitude (21.9%) dominate, followed by temperature (13.3%) and bulk speed (11.7%). Notably, the transverse velocity component Vy and east‐west flow angle together contribute 13.7%, identifying flow deflection as a consistent but previously under‐quantified SIR signature. By producing continuous probabilities rather than binary labels, SIREN enables flexible threshold tuning for operational use and provides a template for compact, interpretable deep‐learning systems in space weather.
Related
- Modulation of Galactic Cosmic Rays in Corotating Interaction Regions: Diffusion Dominates over Drifts and Advection
- A living catalog of stream interaction regions in the Parker Solar Probe era
- Mapping Wolf-Rayet winds with spectro-interferometry I. Detection of multiple co-rotating interaction regions in WR 6
- E-INSPIRE - II. Finding relics from wide-sky multi-band surveys: A proof-of-concept machine learning regression algorithm
- Interpretable human-label-free deep learning for real-bogus classification with uncertainty quantification
BibTeX
@article{mayank2026finding,
title = {{Finding Novel Precursors for Solar Wind Stream Interaction Regions With Interpretable Deep Learning}},
author = {Prateek Mayank and Yogesh and Enrico Camporeale and D. Chakrabarty and Lan K. Jian and Gregory G. Howes and Thomas E. Berger},
journal = {Space Weather},
year = {2026},
month = oct,
doi = {10.1029/2026sw005141},
url = {https://doi.org/10.1029/2026sw005141},
note = {CC BY},
}
RIS
TY - JOUR
TI - Finding Novel Precursors for Solar Wind Stream Interaction Regions With Interpretable Deep Learning
AU - Prateek Mayank
AU - Yogesh
AU - Enrico Camporeale
AU - D. Chakrabarty
AU - Lan K. Jian
AU - Gregory G. Howes
AU - Thomas E. Berger
PY - 2026
DA - 2026/10/28
JO - Space Weather
DO - 10.1029/2026sw005141
UR - https://doi.org/10.1029/2026sw005141
AB - Solar wind stream interaction regions (SIRs) drive recurrent geomagnetic storms, yet most existing catalogs rely on expert inspection and simple thresholds that are subjective and can miss events with complex morphologies. We present SIREN (SIR Encoder Network), a lightweight Transformer‐based model for per‐timestep SIR detection from in situ solar wind observations. The model ingests sequences of 11 solar wind parameters, spanning magnetic field, velocity, and thermodynamic properties. With approximately 100,000 trainable parameters in a two‐layer encoder architecture, SIREN is trained using weighted binary cross‐entropy loss and a cosine annealing learning rate. Platt scaling is applied to produce well‐calibrated detection probabilities. On a held‐out test set of 96 events, the calibrated model achieves a high skill score and analysis of the self‐attention weights confirms that the model concentrates on the SIR, grounding its decisions in the physically relevant portion of each sequence. Integrated Gradients attribution reveals a quantifiable feature hierarchy: proton density (23.4%) and magnetic field magnitude (21.9%) dominate, followed by temperature (13.3%) and bulk speed (11.7%). Notably, the transverse velocity component Vy and east‐west flow angle together contribute 13.7%, identifying flow deflection as a consistent but previously under‐quantified SIR signature. By producing continuous probabilities rather than binary labels, SIREN enables flexible threshold tuning for operational use and provides a template for compact, interpretable deep‐learning systems in space weather.
SN - 1542-7390
ER -