Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Predicting the Emergence of Solar Active Regions Using Machine Learning

Spiridon Kasapis, Irina N. Kitiashvili, Alexander G. Kosovichev, John T. Stefan, Bhairavi Apte

TL;DR

This study addresses the challenge of forecasting solar active region emergence before it becomes visible on the surface by leveraging machine learning on subsurface acoustic-power maps derived from SDO/HMI observations. An 80-hour rolling window of acoustic power across four frequency bands feeds a 3-layer LSTM (64 hidden units) to predict 5-hour-ahead continuum-intensity variations, using 63 tiles per AR and data from 40 ARs. The authors define a pre-emergence signal based on a thresholded mean intensity change and demonstrate that the model can reproduce the initial intensity decrease for at least one emerging AR (AR13179), highlighting the potential for space-weather precursor forecasting even without direct magnetic-field inputs. They also note limitations—particularly late-stage prediction accuracy as ARs evolve—and propose future improvements including incorporating magnetic-field measurements and improved calibration to enhance predictive capability.

Abstract

To create early warning capabilities for upcoming Space Weather disturbances, we have selected a dataset of 61 emerging active regions, which allows us to identify characteristic features in the evolution of acoustic power density to predict continuum intensity emergence. For our study, we have utilized Doppler shift and continuum intensity observations from the Helioseismic and Magnetic Imager (HMI) onboard the Solar Dynamics Observatory (SDO). The local tracking of 30.66 x 30.66-degree patches in the vicinity of active regions allowed us to trace the evolution of active regions starting from the pre-emergence state. We have developed a machine learning model to capture the acoustic power flux density variations associated with upcoming magnetic flux emergence. The trained Long Short-Term Memory (LSTM) model is able to predict 5 hours ahead whether, in a given area of the solar surface, continuum intensity values will decrease. The performed study allows us to investigate the potential of the machine learning approach to predict the emergence of active regions using acoustic power maps as input.

Predicting the Emergence of Solar Active Regions Using Machine Learning

TL;DR

This study addresses the challenge of forecasting solar active region emergence before it becomes visible on the surface by leveraging machine learning on subsurface acoustic-power maps derived from SDO/HMI observations. An 80-hour rolling window of acoustic power across four frequency bands feeds a 3-layer LSTM (64 hidden units) to predict 5-hour-ahead continuum-intensity variations, using 63 tiles per AR and data from 40 ARs. The authors define a pre-emergence signal based on a thresholded mean intensity change and demonstrate that the model can reproduce the initial intensity decrease for at least one emerging AR (AR13179), highlighting the potential for space-weather precursor forecasting even without direct magnetic-field inputs. They also note limitations—particularly late-stage prediction accuracy as ARs evolve—and propose future improvements including incorporating magnetic-field measurements and improved calibration to enhance predictive capability.

Abstract

To create early warning capabilities for upcoming Space Weather disturbances, we have selected a dataset of 61 emerging active regions, which allows us to identify characteristic features in the evolution of acoustic power density to predict continuum intensity emergence. For our study, we have utilized Doppler shift and continuum intensity observations from the Helioseismic and Magnetic Imager (HMI) onboard the Solar Dynamics Observatory (SDO). The local tracking of 30.66 x 30.66-degree patches in the vicinity of active regions allowed us to trace the evolution of active regions starting from the pre-emergence state. We have developed a machine learning model to capture the acoustic power flux density variations associated with upcoming magnetic flux emergence. The trained Long Short-Term Memory (LSTM) model is able to predict 5 hours ahead whether, in a given area of the solar surface, continuum intensity values will decrease. The performed study allows us to investigate the potential of the machine learning approach to predict the emergence of active regions using acoustic power maps as input.
Paper Structure (5 sections, 1 equation, 4 figures)

This paper contains 5 sections, 1 equation, 4 figures.

Table of Contents

  1. Introduction
  2. Data Preparation
  3. Long Short-Term Memory Modeling
  4. Prediction of the Continuum Intensity Variations
  5. Discussion and Conclusions

Figures (4)

  • Figure 1: Scheme of the workflow (top panel) and the machine learning operations (bottom panel) to predict the emergence of active regions. Denoted at $t_p$ is the time of prediction in each time step of the LSTM rolling window approach.
  • Figure 2: The tracked $30.66\times30.66$ degree patch for AR13179 (left) within the HMI continuum intensity full disc and the 9 by 9 grid it was split in (right). The colored patch corresponds to the variations of the acoustic power and continuum intensity shown in the same color in Figure \ref{['fig:acoustic_power_timelines']}.
  • Figure 3: Time evolution of the mean acoustic power for 63 tiles obtained from the $30.66\times30.66$-degrees area for the 3-4 mHz frequency range (top panel). The middle and bottom panels show the evolution of the normalized acoustic power and the continuum intensity corrected for the geometric effect. The blue curve shows the emergence and evolution of AR12085 in the same color-coded middle tile in Figure \ref{['fig:normalization']}.
  • Figure 4: Continuum intensity prediction for AR13179 by the trained LSTM model. The left panels show the predicted (solid curves) and actual (dashed) continuum intensity variations in tiles marked on the right continuum intensity images. The vertical lines indicate the moment half an hour before emergence (blue) and a moment after emergence (orange). The right panels show the whole 30.66 by 30.66 degrees continuum intensity areas corresponding to these moments before (top) and after emergence (bottom image). In the snapshots on the right, enclosed in a red square is a tile that exhibits the formation of an active region (active tile, upper plot) and corresponds to the blue middle tile in Figure \ref{['fig:normalization']}. Enclosed in the green square is a quiet-Sun region (quiet tile, bottom plot). The grey horizontal dashed line indicates the threshold that intensity has to exceed in order to have emergence in this study.