- Scientists expect Surya to extract ideas from the complex magnetic processes of the Sun
- The researchers processed nine years of images of the Solar Dynamics Observatory
- Surya achieved an improvement of 16% in the accuracy of the flag classification
IBM and NASA have introduced Surya, the first open source base model for solar physics.
IBM says that the AI model, whose name comes from the Sanskrit word for the sun, is trained to forecast solar activities, such as flares and storms that can interrupt satellites, navigation systems and electrical networks.
It has been made available through Hugging Face, Github and the IBM Terratorch Library, along with a collection of data set called Suryabench.
From land data to solar forecasts
The project occurs as the dependence on space technology expands, from aviation and communication to future deep space missions.
The prediction of the solar climate remains a difficult task, since these events originate millions of miles in a body whose physics is still partially understood.
“We have been on this trip to boost the limits of technology with NASA since 2023, offering pioneering fundamental models to obtain an unprecedented understanding of our planet Earth,” said Juan Bernabé-Moreno, director of IBM in charge of scientific collaboration with NASA.
“With Surya we have created the first base model to look at the sun in the eye and predict their moods.”
This collaboration follows the previous work of IBM and NASA in models promoted by AI for the prediction of the Earth and the Meteorarate, which led to the development of the Prithvi model that analyzed the satellite data to help the studies of climate and atmospheric systems.
With Surya, they are trying something similar to the sun, turning years of high -resolution solar images of the NASA solar dynamics observatory in a kind of digital twin.
Scientists expect the model to allow forecasts that go beyond whether a flare will occur.
The first reports suggest that Surya can generate visual predictions of flags high resolution up to two hours before they occur, double the delivery time of traditional methods.
That would mean an additional preparation time for astronauts and critical infrastructure operators on Earth.
To build Surya, the researchers processed nine years of images of the Solar Dynamics Observatory, which captures the sun every 12 seconds in multiple wavelengths.
They used a long -term vision transformer with spectral activation to administer the immense data load.
The model was trained not only to analyze the current conditions, but to infer how future observations would be seen, testing their precision against real data.
“We want to give the land as long as possible,” said Andrés Muñoz-Jaramillo, solar physicist at the Southwest Research Institute and the main scientific scientist in the project.
“Our hope is that the model has learned all the critical processes behind the evolution of our star over time so that we can extract processable ideas.”
Like other large language models and AI tools, Surya raises questions about whether its results should be treated as a discovery or increase in human experience.
However, their sponsors emphasize automation and efficiency, pointing out a claimed improvement of 16% in the precision of classification of the Bengal.
Even so, the prognosis remains far from being safe, since the activity of the sun involves many processes that remain little known.
While Surya is described as a step towards a better anticipation of solar threats, researchers are careful not to present it as a final response.
Instead, they frame it as a bridge that can help scientists work with massive data more effectively.
As with any IA or LLM writer, their predictions are limited by the data in which it has been trained and the assumptions integrated in its design.
