Causal Discovery in Photospheric Magnetic Field Time Series for Interpretable Solar Flare Prediction
- 1Utah State University
Abstract
Deep learning has proven to be a vital advancement in many fields that present difficult classification and prediction tasks. This includes astrophysics, particularly space weather prediction, where accurate modeling has become essential for protecting satellite and ground infrastructure from the impacts of solar flares and solar energetic particle events. Despite their strong predictive ability, one issue with deep learning as a method for space weather prediction is the poor interpretability of these “black box” methods. A possible solution to this comes in the form of a structured deep learning framework named Amortized Causal Discovery (ACD), which aims to enforce an easily interpretable layer in the neural network that represents the causal summary graph. We apply this method to the Space Weather Analytics for Solar Flares dataset, a rich source of multivariate time series (MVTS) data of solar active regions, to demonstrate that high-intensity solar flares can be detected ahead of time by comparing their causal summary graphs with those of other flare events. In this work, we use graph metric statistics as well as objective performance tests against state-of-the-art classifiers to demonstrate the depth of insight that ACD can produce into differences between MVTS samples in the context of solar flare prediction.
Related
- Cross-layer coupling in active-region upflows: Statistical connections between photospheric magnetic field, chromospheric activity, and coronal blueshifts
- Broken Power-law Distributions of Photospheric Magnetic Concentrations and Implications for Solar Dynamo
- Discovery of Interpretable Surrogates via Agentic AI: Application to Gravitational Waves
- ARENA: Active Region attENtion Aggregation for interpretable solar flare prediction
- Discovery of a Nearby Ultra-fast Magnetic White Dwarf Propeller: Common Pathways to Magnetic Cataclysmic Variables
BibTeX
@article{nelson2026causal,
title = {{Causal Discovery in Photospheric Magnetic Field Time Series for Interpretable Solar Flare Prediction}},
author = {Nathan Nelson and Onur Vural and MohammadReza EskandariNasab and Shah Muhammad Hamdi and Soukaina Filali Boubrahimi},
journal = {The Astrophysical Journal Supplement Series},
year = {2026},
month = oct,
doi = {10.3847/1538-4365/aea84c},
url = {https://iopscience.iop.org/article/10.3847/1538-4365/aea84c},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Causal Discovery in Photospheric Magnetic Field Time Series for Interpretable Solar Flare Prediction
AU - Nathan Nelson
AU - Onur Vural
AU - MohammadReza EskandariNasab
AU - Shah Muhammad Hamdi
AU - Soukaina Filali Boubrahimi
PY - 2026
DA - 2026/10/22
JO - The Astrophysical Journal Supplement Series
DO - 10.3847/1538-4365/aea84c
UR - https://iopscience.iop.org/article/10.3847/1538-4365/aea84c
AB - Deep learning has proven to be a vital advancement in many fields that present difficult classification and prediction tasks. This includes astrophysics, particularly space weather prediction, where accurate modeling has become essential for protecting satellite and ground infrastructure from the impacts of solar flares and solar energetic particle events. Despite their strong predictive ability, one issue with deep learning as a method for space weather prediction is the poor interpretability of these “black box” methods. A possible solution to this comes in the form of a structured deep learning framework named Amortized Causal Discovery (ACD), which aims to enforce an easily interpretable layer in the neural network that represents the causal summary graph. We apply this method to the Space Weather Analytics for Solar Flares dataset, a rich source of multivariate time series (MVTS) data of solar active regions, to demonstrate that high-intensity solar flares can be detected ahead of time by comparing their causal summary graphs with those of other flare events. In this work, we use graph metric statistics as well as objective performance tests against state-of-the-art classifiers to demonstrate the depth of insight that ACD can produce into differences between MVTS samples in the context of solar flare prediction.
SN - 0067-0049
SN - 1538-4365
ER -