Using Machine Learning To Predict The Trillion Dollar Solar Storm

The biggest Solar Storm ever recorded is the Carrington Event of 1859, which caused telegraph networks around the globe to fail catastrophically, with telegraph paper catching fire and operators receiving electric shocks from equipment.

There have been 26 significant 'space weather' events affecting Earth over the last 50 years. These solar events can severely disrupt the Earth's magnetosphere (the boundary between the Earth's magnetic field and the solar wind), and pose a direct threat to electrical infrastructure - knocking out technologies that we rely on every single day, like GPS signals, electrical grids, computers and satellites. To put it lightly, if a major event were to happen tomorrow, it's likely to cost at least $2 trillion in damages in the first year alone.

So, we're all doomed - right? Well, not exactly.

Luckily for us, NASA has founded a Solar Space Team, which sits within the Frontier Development Lab (FDL). Bringing together the best technological minds from across the world, the FDL tackles space science head-on, deploying specialised machine learning techniques to help protect the planet from space weather. Soon, they hope to be able to accurately predict any future 'trillion dollar storm'.

What exactly is 'Space Weather'?

'Space weather' refers to conditions on the Sun and in the solar wind that influence our planet's magnetosphere and upper atmosphere. Solar flares, filament eruptions, solar radio bursts, active solar regions and enhanced solar wind all play a part in space weather and the behaviour of the magnetosphere.

How much of a threat is it really?

While many space weather events pass us by without much fuss, the main source of dangerous space weather is the violent and sudden release of magnetised plasma bubbles from our Sun, or Coronal Mass Ejections (CMEs).

Occurring anywhere between several times a day to once a week, CMEs send shock waves and bursts of energetic particles that stream near the speed of light straight towards Earth. These blasts of radiation make it difficult for satellite signals to penetrate the atmosphere, interrupting vital services including GPS and the WAAS system used by air traffic. They do this by destabilising our planet's ionosphere - the part of the atmosphere through which radio signals travel, and through which many satellites orbit.

According to NASA, we have more than a 1 in 10 chance of being hit by a CME sometime in the next decade. In fact, Earth is no stranger to these events, with the most recent happening only 15 years ago. Dubbed the 'Halloween Solar Storms of 2003', numerous satellites and communications systems were affected, there was an hour-long power outage across Sweden, and aircrafts were advised to avoid certain altitudes near the Polar Regions.

However, as bad as that was, the biggest Solar Storm ever recorded is the Carrington Event of 1859, which caused telegraph networks around the globe to fail catastrophically, with telegraph paper catching fire and operators receiving electric shocks from equipment. Whilst that was deemed catastrophic back then, many believe that in today's world of interconnectivity the effects would be much, much worse.

How is Machine Learning helping?

For the first time, sophisticated data processing and machine learning is enabling scientists to make significant breakthroughs in defending our planet from asteroids and solar weather, and in particular, monitoring CMEs. Machine learning can now be used to detect early warning signs of potentially hazardous solar storms, which is improving predictive models of major solar events and the emergence of new sunspot groups that predict the state of the Sun tomorrow.

Machine learning also offers the opportunity to analyse variations in the solar magnetic field and solar corona using data from the Solar Dynamics Observatory (SDO, surface vector magnetograms and EUV images). This data can be used to discover relationships between the observed magnetic activity in the photosphere and corona, and to identify the agents that drive solar eruptive events, such as flares and CMEs.

The scale of the challenge is not to be underestimated, with each data centre recording a reading for every minute of every day, amounting to 525,600 readings per data centre, per year. Our solar team has examined this across 14 data centres from 2010 to 2016, meaning that in a non-leap year there are 51,508,800 records to train and validate neural networks on. This is where the latest data processing techniques come into their own, with teams currently deploying complex neural networks to increase their predictive capabilities.

With the ability to effectively crunch and make predictions on such large datasets, other potential breakthroughs become increasingly possible. Amongst others, scientists are already looking to such technology to locate and model the orbits of comets, and to change radar images of asteroids into accurate 3D models to help determine shape and spin.

To infinity, and beyond!

There is no doubt that in the future, the pairing of data aggregation with machine learning techniques will become the critical enabler of future Space Exploration and defending our planet from asteroids and solar weather. The importance of this cannot be underestimated, as space weather events can affect our ground-based technological systems, and in serious cases, can even endanger human life or health. Monitoring space weather will also allow for us to move out of LEO (Launch and Early Orbit) and into deep space exploration; permanently manning facilities on the moon and fulfilling NASA's goal to visit Mars within the next two decades.

The space sector, and engagement with space, is changing rapidly, with the global space market expected to grow to be worth $400bn by 2030. As this 'New Space' era generates increasing amounts of data in years to come, it is even more important for the technology community to provide new innovations to help us to succeed in tackling this global issue.