NASA and IBM built an AI to predict solar flares before they hit Earth

An artificial intelligence model trained on NASA satellite imagery can forecast what the sun will look like hours into the future – even predicting the appearance of solar flares that may warn of dangerous space weather for Earth.

“I love to think of this model as an AI telescope where you can look at the sun and you can understand the moods,” says Juan Bernabé-Moreno at IBM Research Europe.

The sun’s moods matter because outbursts of solar activity can bombard Earth with high-energy particles, X-rays and extreme ultraviolet radiation. These can disrupt GPS and communications satellites, and potentially harm astronauts and even people on commercial airlines. Solar flares can be followed by coronal mass ejections, which may disrupt Earth’s own magnetic field and create geomagnetic storms capable of knocking out power grids.

Bernabé-Moreno and his colleagues at IBM and NASA trained an AI model called Surya, after the Sanskrit word for sun, on nine years of data from NASA’s Solar Dynamics Observatory. The satellite captures ultra-high-resolution images of the sun in 13 different wavelengths. The AI model learned to identify patterns in the visual data and generate images of what the sun would look like from the observatory’s point of view in the future.

When tested on historical solar flare data, the Surya model predicted the occurrence of a solar flare within the next day with 16 per cent better accuracy than a standard machine learning model. It could also generate the visual image of a flare the observatory would see up to two hours in the future.

“The power of AI is that it has the ability to learn the physics in a more roundabout way – it kind of develops an intuition for how the physics works,” says Lisa Upton at Southwest Research Institute in Colorado.

Upton says she is especially interested in whether the Surya model can help predict solar activity on the far side of the sun and at the poles, where NASA’s scientific instruments can’t make direct observations. Surya does not explicitly attempt to model the far side of the sun, but it has still proven successful in predicting how the sun will look several hours in the future, when part of the far side has rotated into view, says Bernabé-Moreno.

But it is unclear if the AI model can address existing challenges in predicting exactly how solar activity may impact Earth, says Bernard Jackson at the University of California, San Diego. That is because there is currently no way to directly observe the magnetic field configurations between the sun and Earth, which is what determines the paths of the high-energy particles travelling outward from our star.

Bernabé-Moreno says the model is currently intended for use by scientists, but future integrations with other AI systems that can harness Surya’s capabilities to answer basic questions about future solar activity might make it more accessible to power grid operators or satellite constellation owners as part of an early warning system.