Solar Flare Prediction Advances with Data-Driven Techniques and Machine Learning Paradigms

Solar flares represent a significant challenge for space weather forecasting, with the potential to disrupt satellites, communications and power grids on Earth. Mingfu Shao, Suo Liu, and Haiqing Xu, along with their colleagues, comprehensively review the latest advances in predicting these powerful eruptions from the Sun. Their work traces the evolution of forecasting techniques, from early statistical methods to cutting-edge machine learning and deep learning approaches, including the recent emergence of Multimodal Large Models. By evaluating the performance of current operational flare forecasting platforms, the team identifies critical limitations and provides valuable insights to guide future improvements in system design and optimisation, ultimately enhancing our ability to protect vital technologies in space and on the ground.

Solar Flare Prediction, From Patterns to Models

Researchers are transitioning from traditional methods to advanced machine learning techniques, particularly foundation models, to improve solar flare forecasting. Early approaches relied on classifying active regions based on sunspot morphology and magnetic field properties, but these methods often struggled with accuracy and required significant expertise. Now, deep learning models, including Long Short-Term Memory networks, Convolutional Neural Networks, and Transformers, are being employed to analyse data from missions like the Solar Dynamics Observatory and the Advanced Space-based Solar Observatory. Combining multiple deep learning models and creating interpretable models are also being explored.

Inspired by large language models, scientists are developing foundation models for heliophysics, pre-training them on vast amounts of solar data and fine-tuning them for tasks like flare prediction. Key examples include Surya, a model specifically designed for heliophysics, and models trained on data from the Solar Dynamics Observatory. Researchers are also investigating reinforcement learning and developing multimodal models that can process images and time series data. A growing movement towards open-source foundation models is fostering collaboration and accelerating research, promising improved space weather prediction and mitigation of impacts on Earth and technology in space.

Long-term Flare Prediction Using Magnetic Fields and X-rays

This study pioneers a comprehensive approach to solar flare prediction, leveraging data from ground-based telescopes and space-based satellites. Researchers analysed nearly four solar cycles of vector magnetic field observations collected by a telescope in China, establishing a crucial historical dataset. Complementing these observations, the team incorporated data from the GOES satellite, examining X-ray flux measured at one-minute intervals. They meticulously examined data from the GOES-18 and GOES-19X-Ray Sensors, classifying flare levels based on peak flux. Researchers also utilised data from the Solar and Heliospheric Observatory, the Solar Dynamics Observatory, and the Advanced Space-based Solar Observatory, providing extensive data on the solar magnetic field and active regions. The Michelson Doppler Imager and the Helioseismic and Magnetic Imager delivered comprehensive and high-resolution magnetograms. While acknowledging limitations in GOES data, the team emphasises its continued indispensability when integrated with other observational datasets.

Surya Model Forecasts Solar Flares Accurately

Recent work demonstrates substantial progress in solar flare prediction, driven by advancements in data processing and sophisticated machine learning techniques. Researchers have successfully employed multimodal large models, achieving significant improvements in forecasting accuracy and paving the way for a “one model, multiple tasks” framework in heliophysics. The Surya model, pretrained on high-resolution solar images, attained a True Skill Statistic of 0. 436, exceeding the performance of established baselines like AlexNet and ResNet50. A comprehensive assessment of operational flare forecasting systems, including DeepSun, DeepFlareNet, SolarFlareNet, MViT, and the NOAA/CMCC system, reveals varying levels of performance.

SolarFlareNet, a Transformer-based model, achieved a True Skill Statistic exceeding 0. 83 for 24-hour forecasts of ≥C-class flares, surpassing traditional models. Further advancements were made with the MViT model, achieving a True Skill Statistic of 0. 74 for ≥M-class flare prediction. The study highlights the increasing sophistication of data analysis applied to solar flare prediction, driven by advancements in space observation and data processing capabilities. The investigation details the historical progression of data sources used in flare prediction, beginning with ground-based telescope observations and transitioning to the more comprehensive data provided by space-based satellites, such as the GOES series. While ground-based data provided crucial early statistical information, its use is limited by factors like instrument age and atmospheric conditions. Further improvements in data quality and model development are needed to enhance the accuracy and reliability of operational flare forecasting systems.