Suppose a U.S. Navy destroyer is operating in the Western Pacific and is impacted by the effects of a massive solar flare that has caused a ripple effect across the globe, impacting power grids, destroying electrical infrastructure, and causing havoc to electronic communications. What the ship’s captain needs to know is how long this will last?
If Jeff Reep, an astrophysicist at the U.S. Naval Research Laboratory has his way, the captain would be able to answer that question with real-time forecasting. Reep has developed a method using a machine learning algorithm that would allow some prediction to the fleet for how long a flare might actually last.
“I study solar flares and I spent my career modeling them,” Reep said. “We generally think of flares in terms of their size or their brightness, but we often ignore the part about their duration.”
Solar flares, from the Navy's perspective, impact operations due to disturbances in the ionosphere, home to virtually all the charged particles in Earth's atmosphere right at the edge of space, forming the boundary between Earth's lower atmosphere, where we live and breathe, and the vacuum of space. Ionospheric disturbances change the propagation of radio signals in the upper atmosphere, an effect that is familiar to radio operators around the world.
What makes predicting the duration difficult is that the length of the flare does not relate to the size of a flare. In fact, there is no correlation to any of the basic parameters of a flare. For Reep and others in his field, modeling solar flares involves computationally intensive models that cannot be run in real time.
Reep has been looking at solar flare durations for a few years from the perspective of physical modeling. Thanks to a NASA “Living With a Star” grant, Reep combined the expertise of NRL solar physicists and the expertise of an NRL ionosphere modeling group to investigate the effects that a solar flare produces. This work is aimed at predicting what the impact on the ionosphere would be, and consequently, what the impact on Naval operations would be.
“When we create a model of the solar flare, it might take weeks of computational time to run,” Reep said. “But a flare might be done in an hour, which means it can’t be used for real-time predictions. So we wanted something that would give us a quicker answer that the Navy could actually use.”
In comes the Random Forest Regressor, a machine learning technique that is fed a series of variables and that uses “decision trees” to predict features of an event. Think of it in terms of the game show, “Who Wants to be a Millionaire?” On the show, contestants have three lifelines, one of which is to poll an audience of 100 people. If no one knows which of the four answers is correct, the answer splits randomly. So 25 in each column. But if 10 people in the audience know the answer, then 10 people are going to be in one of those four columns definitively, and the other 90 people will be split randomly.
“Having that small number of people get the question right makes the answer pretty obvious because that one answer will stick out above the rest,” Reep said. “Each individual decision tree doesn't have to be good at making predictions, but when you throw 1,000 of them together, if a small fraction of them are good at making predictions, you can actually get something that's reasonable.”
So how close are researchers to providing the answer to the captain of the ship on how long the flare-effects will last? “The challenge is acquiring all the data in real time so we can have a functional model that's constantly running, and when a flare goes off, it can make a real-time prediction," said Reep. "We still have a few roadblocks to get past.”