Forecasting Continuum Intensity for Solar Active Region Emergence Prediction using Transformers

2026

Read Forecasting Continuum Intensity for Solar Active Region Emergence Prediction using Transformers (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-13144.txt.

Introduction

A team study, with no named lead author in the record, seeks earlier flags for new active zones on the Sun. Such zones dim the light from the Sun as they rise. To call that dim in time is key for space weather work.

Past tools used LSTM nets, a kind of memory net, to guess the drop in bright flux as a zone nears. The new work widens the set of net forms tried. It also widens the marks that the nets must guess. The shift to newer forms, named in the title as Transformers, aims for flags that come sooner and hold up better. Sun teams gain more lead time to guard craft and grids.

Catalog record

YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE68 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-13144

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Forecasting Continuum Intensity for Solar Active Region Emergence Prediction using Transformers Abstract Early and accurate prediction of solar active region (AR) emergence is crucial for space weather forecasting. Building on established Long Short-Term Memory (LSTM) based approaches for forecasting the continuum intensity decrease associated with AR emergence, this work expands the modeling with new architectures and targets. We investigate a sliding-window Transformer architecture to forecast continuum intensity evolution up to 12 hours ahead using data from 46 ARs observed by SDO/HMI. We conduct a systematic ablation study to evaluate two key components: (1) the inclusion of a temporal 1D convolutional (Conv1D) front-end and (2) a novel ‘Early Detection’ architecture featuring attention biases and a timing-aware loss function. Our best-performing model, combining the Early Detection architecture without the Conv1D layer, achieved a Root Mean Square Error (RMSE) of 0.1189 (representing a 10.6% improvement over the LSTM baseline) and an average advance warning time of 4.73 hours (timing difference of -4.73h), even under a stricter emergence criterion than previous studies. While the Transformer demonstrates superior aggregate timing and accuracy, we note that this high-sensitivity detection comes with increased variance compared to smoother baseline models. However, this volatility is a necessary trade-off for operational warning systems: the model’s ability to detect micro-changes in precursor signals enables significantly earlier detection, outweighing the cost of increased noise. Our results demonstrate that Transformer architectures modified with early detection biases, when used without temporal smoothing layers, provide a high-sensitivity alternative for forecasting AR emergence that prioritizes advance warning over statistical smoothness. † † journal: JGR: Machine Learning and Computation † † authors: Jonas Tirona, Sarang Patil, Spiridon Kasapis, Eren Dogan, John Stefan, Irina N. Kitiashvili, Alexander G. Kosovichev, Mengjia Xu New Jersey Institute of Technology, Newark, NJ, USA Princeton University, Princeton, NJ, USA NASA Ames Research Center, Moffett Field, CA, USA Plain Language Summary Solar active regions are areas of intense magnetic flux on the Sun that drive space weather events, such as solar flares, which can disrupt technologies on Earth. Forecasting the emergence of these regions is challenging because the warning signs, subtle variations in the Sun’s acoustic and magnetic patterns, are faint and difficult to detect before the region becomes visible. In this study, we applied a Transformer machine learning architecture, designed to recognize complex sequential patterns, to analyze these precursor signals to predict the emergence of solar active regions. We compared standard models against a specialized ”Early Detection” architecture. We found that standard architectures often prioritize statistical stability, and the addition of convolutional layers effectively smooth out the faint data spikes that signal the start of an event, causing inaccuracies in timing prediction. By modifying the model to be more sensitive to these initial fluctuations, our approach successfully predicted the emergence of active regions approximately 5 hours in advance. This demonstrates that for operational space weather forecasting, models must be designed to prioritize early detection of faint signals rather than prioritizing the smoothness of the prediction. 1 Introduction Solar active regions (ARs) manifest from the Sun’s magnetic field emerging from the convection zone to the photosphere ( 26 , e.g.,) , from where they drive events like space weather conditions by producing space weather events such as solar flares ( 23 , e.g.,) , coronal mass ejections ( 8 , CMEs; e.g.,) and Solar Energetic Particles ( 28 , SEPs; e.g.,) . These events, if sufficiently intense, can inject energetic particles in Earth’s magnetosphere and affect technological infrastructure ( 7 ; 13 ; 18 , e.g.) . Therefore, understanding when ARs emerge on the photosphere and produce flares, CMEs, and SEPs is crucial for predicting such geoeffective space-weather events and mitigating their effects on space-based infrastructure. The emergence process of ARs involves rising magnetic flux through the solar interior, producing observable signatures in helioseismic data hours to days before becoming visible as sunspots ( 15 ; 9 ; 25 ) . A number of helioseismic studies have confirmed that variations in the travel times of interior-penetrating acoustic waves can signal the upcoming flux emergence that later will form in ARs, before it can be observed visually on the photosphere ( 2 ; 17 ; 1 ; 25 , e.g.,) . These travel time variations are extracted through Fourier and cross-correlation analyses of the surface velocity field, and it has been recently shown that direct use of the velocity field—without the use of computationally-expensive helioseismic inversion techniques—can also be effective in retrieving information about subsurface emerging magnetic flux ( 12 ) , ( 14 ) . Recently, advances in machine learning and helioseismology have shown promise for predicting AR emergence using time-series analysis of solar oscillations observed on the solar surface ( 16 ) . An initial architecture for studying such temporal sequence modeling was the Long Short-Term Memory (LSTM) network ( 11 ) , originally introduced by 5 . Designed to mitigate the vanishing gradient problem inherent in standard Recurrent Neural Networks (RNNs), LSTMs utilize a gating mechanism to regulate the flow of information, making them particularly effective for learning order-dependence in time-series data. 11 ; 12 demonstrated that LSTMs could successfully forecast the continuum intensity associated with AR emergence by analyzing acoustic power and magnetic flux variations derived from the Helioseismic and Magnetic Imager (HMI) instrument data ( 22 ; 6 ) onboard the Solar Dynamics Observatory (SDO) ( 21 ) , achieving predictions 10-29 hours in advance for several test cases. However, LSTMs are not immune to the limitations of recurrent architectures. Their sequential processing nature requires information to traverse every intermediate time step, which can cause the backpropagated error gradients—the signals used to update model weights—to diminish over extended sequences. This attenuation effectively prevents the model from learning relationships between distant time points. To address these limitations and capture complex temporal dynamics more effectively, Transformer architectures ( 27 ) have revolutionized sequence modeling across multiple domains through their self-attention mechanisms, which enable direct modeling of relationships between any two positions in a sequence regardless of their temporal distance. In time-series forecasting applications, Transformers have demonstrated superior performance over recurrent models by effectively capturing both local patterns and global dependencies ( 30 ; 29 , e.g.,) . For solar physics applications, the ability to simultaneously analyze multiple temporal scales may prove particularly valuable given the multi-scale nature of the Sun’s magnetic activity. This work extends the developments of predicting AR emergence by investigating the optimal architecture for a sliding-window Transformer. We present the first application of Transformer architectures to predict continuum intensity evolution driven by active region emergence. We conduct a systematic ablation study to evaluate two key components of our proposed forecasting framework: (1) the inclusion of a temporal 1D convolutional (Conv1D) front-end, and (2) a novel ‘Early Detection’ architecture featuring specialized attention biases and a timing-aware loss function, alongside a detailed analysis of model behavior during the emergence phase. This paper is organized as follows: Section 2 describes the dataset, preprocessing procedures, and the architectures of our experimental setup alongside the baseline LSTM model, detailing the loss functions and optimization strategies employed. Section 3 presents quantitative performance metrics, hyperparameter search results, and detailed evaluation within emergence windows, supported by representative diagnostic plots. Section 4 interprets our findings, examining the trade-offs

Read the complete text