TLDR: Surya is a new 366-million-parameter AI foundation model for heliophysics, trained on 15 years of high-resolution Solar Dynamics Observatory (SDO) data. It learns general solar representations, enabling accurate zero-shot forecasting of solar dynamics and flares, and excels in downstream tasks like solar wind prediction, active region segmentation, and EUV spectra forecasting, marking a significant step towards more versatile space weather prediction.
Our Sun, a dynamic star, constantly influences Earth and the entire solar system through phenomena like solar flares and coronal mass ejections. Understanding and predicting these “space weather” events is crucial for safeguarding our technology, from satellites and GPS to power grids and aviation. For decades, scientists have relied on high-resolution observations from missions like the Solar Dynamics Observatory (SDO), but most predictive models have been limited by being task-specific and requiring a lot of labeled data.
Enter Surya, a groundbreaking new foundation model for heliophysics. Developed by a collaborative team of researchers from institutions including NASA Marshall Space Flight Center and IBM Research, Surya is designed to learn general-purpose representations of the Sun from a vast amount of SDO observations. This includes data from eight Atmospheric Imaging Assembly (AIA) channels, which capture extreme ultraviolet (EUV) and ultraviolet (UV) light, and five Helioseismic and Magnetic Imager (HMI) products, which provide magnetic field and velocity measurements. The model is quite large, boasting 366 million parameters, allowing it to process and understand complex solar data.
How Surya Works
Surya employs a sophisticated architecture known as a spatiotemporal transformer. This design incorporates unique features like spectral gating, which helps filter and emphasize important frequency components in the solar data, and long-short range attention, enabling the model to capture both fine-scale details and broad, global patterns of solar activity. The model was initially trained on a “pretext task” – essentially, learning to forecast future solar images. This pre-training helps Surya develop a deep understanding of how the Sun evolves over time. Further optimization was achieved through a technique called autoregressive rollout tuning, where the model learns to predict multiple steps into the future.
Impressive Capabilities
One of Surya’s most exciting features is its “zero-shot” evaluation capability. This means it can forecast solar dynamics and even predict solar flare events without needing specific, labeled training for those tasks. For instance, when tested on past solar flare events, Surya accurately predicted the visual evolution of these powerful eruptions, with its predicted emissions closely matching actual observations. This suggests that the model isn’t just memorizing patterns but is learning the underlying physics of solar evolution.
Beyond zero-shot forecasting, Surya can be efficiently adapted for various specific “downstream” applications using a technique called Low-Rank Adaptation (LoRA). This fine-tuning process allows the model to perform strongly on tasks such as:
Solar Wind Forecasting: Predicting the speed of the solar wind, which is vital for understanding its impact on Earth. Surya significantly outperforms previous models in this area.
Active Region Segmentation: Accurately identifying and outlining active regions on the Sun, which are crucial for predicting flares and coronal mass ejections. Surya achieved higher segmentation quality compared to traditional methods.
Solar Flare Forecasting: Predicting whether a significant flare will occur within a 24-hour window. Surya demonstrated superior performance across key evaluation metrics compared to other deep learning models.
EUV Spectra Prediction: Estimating solar extreme ultraviolet irradiance across 1343 spectral bands, which directly affects satellite functionality and communication systems. Surya’s predictions are more comprehensive and accurate than existing models.
Also Read:
- AI Interprets Weather Diagrams for Precipitation Forecasts
- RadarQA: Enhancing Weather Forecast Evaluation with Multi-modal AI
A Step Forward for Heliophysics
Surya represents a significant advancement in heliophysics. By learning generalized, physics-aware representations from a wealth of high-resolution, multi-instrument solar data, it addresses limitations of older, task-specific models. Its versatility and adaptability mean it can be fine-tuned for diverse applications with minimal additional effort, improving performance, especially for rare but impactful events. While the model, being deterministic, might show some blurring of sharp features in its outputs, it also has a positive denoising effect on certain low-signal channels, making the data more useful for analysis.
The development of Surya establishes a unifying framework for solar forecasting and scientific discovery. The model and datasets are publicly available on Huggingface, and the code for the model and its downstream tasks can be found on GitHub. This openness encourages further research and application in the field. For more detailed information, you can refer to the full research paper: Surya: Foundation Model for Heliophysics.
