- The paper introduces Solaris, a novel foundation model that employs a large Swin Transformer architecture to predict solar atmospheric phenomena.
- It leverages 13 years of full-disk, multi-wavelength solar images with rigorous preprocessing to capture detailed spatiotemporal dynamics.
- Experimental results demonstrate Solaris's superior performance in forecasting solar flares and other dynamic events, paving the way for advanced space weather prediction.
Solaris: A Foundation Model of the Sun
The paper introduces Solaris, the pioneering effort to employ foundation model architecture for forecasting solar atmospheric activities. Leveraging a large Swin Transformer architecture, Solaris represents a significant step toward advancing the domain of solar physics, specifically in modeling and prediction.
Overview of Solaris
Solaris has been built using a comprehensive dataset spanning 13 years of full-disk, multi-wavelength images from NASA's Solar Dynamics Observatory (SDO), thereby encapsulating a complete solar cycle. The model has 109 million parameters, indicative of its significant capacity to handle the complex and dynamic solar data. The architecture is grounded in a foundation model approach, which has been proven effective across various scientific domains, such as biomolecule structure prediction and chemical analysis.
Data and Architecture
The dataset is curated from the Atmospheric Imaging Assembly (AIA), Helioseismic and Magnetic Imager (HMI), and Extreme Ultraviolet Variability Explorer (EVE) instruments. This data was pre-processed meticulously to address sensor deterioration, exposure time variance, and to maintain image quality. With such preprocessing, Solaris operates on uniformly scaled input data, comprising 512x512-pixel images, and includes intense data augmentation and normalization techniques to support model training and inference.
The model's architecture closely follows the established patterns demonstrated by the Aurora model used for atmospheric forecasting. It integrates a perceiver-based encoder, a 3D Vision Transformer (ViT) processor, and a decoder, each specializing in different transformational processes to transmute solar imagery into insightful forecasts. The Swin Transformer backbone ensures the model captures spatiotemporal dynamics efficiently.
Experimental Results
The experimental setup and results underscore Solaris's potent ability to forecast the solar atmosphere, demonstrating a notable understanding of solar dynamics, including latent periods and active solar phenomena such as flares and winds. The model shows competence in pre-training and fine-tuning tasks, significantly outperforming models trained from scratch in low-data scenarios, particularly for the sparsely recorded 1700 Å wavelength.
Implications and Future Directions
Evidently, Solaris represents a notable development in the field of heliophysics, facilitating enhanced comprehension and prediction of the solar atmosphere's behavior. Given the sheer volume of data from the SDO, Solaris can be foundational in tackling broader datasets or even higher-resolution predictions for more granular solar events. Additionally, these results imply a mature capability for the model to engage in downstream tasks post fine-tuning, such as specific wavelength predictions or irradiance forecasting.
Future work may focus on expanding the data spectrum Solaris handles, possibly integrating different SDO data types to escalate predictive accuracy further. Another intriguing direction could involve attempting to quantify Solaris’s understanding of the solar state, which although internally modeled, is not explicitly decoded presently. By scaling up Solaris, prospects open for enriched multi-scale solar dynamics understanding, potentially impacting practical space weather prediction and handling of technological risks on Earth.
The findings presented in this paper set the stage for continued exploration and refinement of foundation models within heliophysics, establishing a baseline upon which succeeding developments can build and diversify.