AI to predict sun's next attack on Earth

JUST before noon on 1 September 1859, an English solar astronomer named Richard Carrington witnessed the biggest solar flare ever recorded. About 18 hours later, an intense magnetic storm hit Earth. Currents induced in telegraph wires in Europe and North America sparked fires.

If the 1859 event were to occur today, it could devastate our modern technological infrastructure. So researchers are now turning to automated image-processing and artificial intelligence to better forecast the sun's behaviour and give us time to prepare for a solar onslaught.

Over the past two decades, several solar flares and magnetic storms of varying intensity have hit Earth. Solar flares are surges of X-rays, gamma rays and extreme ultraviolet radiation, and they can damage electric grids, fry satellite electronics and endanger astronauts in space. Even passengers and pilots on aircraft flying over the poles are at risk. Coronal mass ejections (CMEs), which cause magnetic storms, can strike even closer to home (see "Shock wave blackout").

With advance warning, satellite operators can switch off sensitive high-voltage electronics on their craft, astronauts can avoid space walks and even hide behind radiation shields, and planes can avoid polar routes. Solar observatories that study the sun continuously should be able to give us some warning before an impending storm. But the copious data streaming from these telescopes is extremely difficult to analyse.

That is why Piet Martens of Montana State University in Bozeman and his team are automating the process of studying the sun. The team is focusing its efforts on data from NASA's Solar Dynamics Observatory (SDO), which was launched on 11 February 2010 and is now orbiting 36,000 kilometres above the Earth, in an orbit synchronised with the sun.

The craft takes images of the sun's surface and atmosphere in 10 different wavelengths. It sends back one set of images every 12 seconds. "You need to be able to identify everything you need to inside those 12 seconds," says astronomer James McAteer at New Mexico State University in Las Cruces. "Otherwise you get backlogged and you are never going to catch up." This mountain of data adds up to a staggering 1.5 terabytes a day.

Besides the challenge of keeping up with the data stream, identifying features on the sun's surface is extremely difficult. "The sun is a challenging subject for automated image analysis," says Erwin Verwichte of the University of Warwick in Coventry, UK. "The solar atmosphere is transparent so that various features appear superimposed within the line of sight, confusing the picture."

So Martens, McAteer and their colleagues have developed 15 programs that use image-processing techniques such as contour or edge recognition to automatically identify features on the sun's surface (arxiv.org/abs/1109.6922). Each program is looking for a different aspect of solar activity. This include flares and CMEs, as well as other features that might indicate that flares or eruptions are imminent, such as filaments, which are bundles of plasma held down by magnetic field lines, coronal loops and sunspots.

The results could give insights into aspects of solar physics, such as the solar cycle. This periodic change in the amount of radiation reaching Earth lasts roughly 11 years, but is highly erratic. The research will also allow astronomers to study the solar surface in detail and note the features that precede each event.

For now, the researchers have created dedicated programs for each feature they want to study. "This is not the way to go in the long run," says Martens.

To make the process generic, his team is using techniques developed to identify breast tumours. This involves splitting a 1.6-million-pixel image into 1024 blocks. For each block, the software calculates the values for various mathematical parameters, such as the entropy (a measure of the chaos in the image). This turns the image into a series of numbers. In breast imaging, this technique highlights regions of breast tissue with specific values that are known to be characteristic of tumours. Martens's team is doing this with SDO images, training the software to learn the defining characteristics of sunspots, filaments and other solar features.

This software can also be used on as-yet-undiscovered features. A new "signature" can be checked against archive images to see if it has ever shown up before, then used as a reference point for future events. With such data, McAteer thinks that solar physicists will finally be able to do high-calibre empirical science.

The techniques will become more important as bigger solar observatories come online, such as the Advanced Technology Solar Telescope to be built in Hawaii later this decade, and the European Space Agency's Solar Orbiter, due to be launched in 2017.

And for us on the ground using GPS devices and living under electricity cables, accurate forecasts of the sun's fiery outbursts cannot come too soon.