- Scientists hope that Surya extracts insights from the complex magnetic processes of the sun
- Researchers processed nine years of images from Solar Dynamics Observatory
- Surya obtained a reported 16% improvement in flare classification accuracy
IBM and NASA have introduced Surya, the first open source foundation model for solar physics.
IBM says that the AI model, whose name comes from the Sanskrit word for the sun, is trained to predict solar activity such as torches and storms that can interfere with satellites, navigation systems and power networks.
It has been made available through Hugging Face, GitHub and IBM’s Terratorch Library along with a data set collection called Suryabench.
From ground data to solar forecasts
The project comes as dependence on space -based technology is expanded, from aviation and communication to future deep space missions.
Predicting sun weather remains a difficult task, given that these events stem from millions of miles away on a body whose physique is still only partially understood.
“We have been on this journey pushing the limits of technology with NASA since 2023 and delivering groundbreaking basic AI models to gain an unprecedented understanding of our planet Earth,” said Juan Bernabé-Moreno, IBM director in charge of scientific cooperation with NASA.
“With Surya, we have created the first foundation model to see the sun in the eye and predict its moods.”
This collaboration follows previous work by IBM and NASA on AI-run models of soil and weather prediction, which led to the development of the prithvi model, analyzing satellite data to help studies of climate and atmospheric systems.
With Surya, they try something similar for the sun that transforms years with high resolution sun image from NASA’s Solar Dynamics Observatory into a kind of digital twin.
Researchers hope the model allows forecasts that go beyond whether a flare will happen.
Early reports suggest that Surya can generate visual predictions with high resolution of torches up to two hours before they occur, which doubles the management time of traditional methods.
That would mean additional preparation time for astronauts and operators of critical infrastructure on Earth.
To build Surya, researchers treated nine years of imagery from Solar Dynamics Observatory, which catches the sun every 12 seconds at several wavelengths.
They used a long-card vision transformer with spectral port to control the huge data load.
The model was trained not only to analyze current conditions, but to derive what future observations would look like, testing its accuracy against real data.
“We want to give the Earth the longest math,” said Andrés Muñoz-Jaramillo, a solar physicist at the Southwest Research Institute and a leading scientist on the project.
“Our hope is that the model has learned all the critical processes behind our star’s development over time so that we can extract actionable insights.”
Like other large language models and AI tools, Surya raises questions about whether its output should be treated as a discovery or as an increase in human expertise.
However, its backers emphasize automation and efficiency and point to an allegedly improvement of 16% in flare classification accuracy.
Still, forecasts remain far from determined as the sun’s activity involves many processes that remain poorly understood.
While Surya is described as a step towards better expectation of solar threats, scientists are careful not to present it as a final answer.
Instead, they frame it as a bridge that can help scientists work with massive data more effectively.
As with any AI author or LLM, its predictions are limited by the data it has been trained on and the assumptions built into its design.
