Machine-learning algorithms may in future be able to warn scientists of harmful radiation storms days in advance of their formation in Earth’s Van Allen belts.
Charged particles from the Sun get trapped in our planet's magnetic field to form two donut-shaped lobes known as Van Allen belts. Satellites traveling through these regions have to be radiation hardened to withstand the constant bombardment of subatomic particles.
Space is unpredictable, and as such, sometimes these particles get whipped into a frenzy by the solar wind, stirring up a flurry of electrons with higher energies than normal. Sensitive equipment, such as cameras, sensors, and antennae, on spacecraft whizzing through these belts can fail when are pelted with these super-energetic particles. Astronauts in orbit can also be hit with higher doses of radiation if they get caught in a belt's storms.
Obviously, you would want to avoid that by maneuvering craft out of the way, or powering down electronics, or pointing gear in another direction, or just knowing things are about to go pear-shaped, and so forth. Thus, an advanced early-warning system would be handy.
Fear not. A trio of boffins at Los Alamos – the American lab best known for designing nuclear weapons in the Second World War – hope to build AI-powered technology that can alert scientists before these freak space weather events take place. To do this, the team developed various machine-learning models, and trained them on readings from National Oceanic and Atmospheric Administration (NOAA) satellites monitoring the radiation in the Van Allen belts.
The software was taught to take into account the speed of the solar wind hitting the belts as well as the energies of subatomic particles within. The algorithms thus attempt to predict from satellite readings whether high concentrations of electrons with energies over a mega-electronvolt will form within the belts over the next two days.
A proof of concept
The researchers split the satellite data, gathered over a period of 42 months, into training, validation, and testing sets. Despite experimenting with models with more complicated architectures, such as convolutional neural networks and long-short term memory networks, it was a simple logistic regression algorithm that prevailed.
When the software was tested on NOAA belt readings, it had a performance of about 0.864. That's not too bad given that 1.000 is a perfect score. But when the code was tested on data from the Los Alamos GEO satellite, which it wasn't trained on, that number plummeted to 0.333, suggesting the model’s accuracy is highly dependent on the training data, and it doesn’t generalize well to patterns it hasn’t seen before.
The results were published in a paper in Space Weather. “With the expectation that similar patterns may reveal themselves in the future, our model is capable of making predictions by capturing some critical signatures as a precursor to those future events," said Youzuo Lin, co-author of the paper and a computational scientist at Los Alamos.
There is good reason for building a predictive model now more than ever, Yue Chen, co-author of the paper and a space scientist at Los Alamos, told The Register on Tuesday:
“Given that the Van Allen Probes, which provided important data about space weather, recently de-orbited, we no longer have direct measurements about what’s happening in the outer electron radiation belt. Our new model uses existing data sets to ‘learn’ patterns and predict future storms so satellite operators can take protective measures, including temporarily shutting down part of or even the whole satellite to avoid damage.”
We also asked Chen why the software performed so poorly with GEO satellite data versus the NOAA readings, and the scientist told us:
"To put it in a simple way, we believe that the dynamics of electrons at GEO is controlled by very different physics compared to electrons at smaller L-shells. In other words, predictions of electrons at GEO can be further improved by involving more input parameter(s) to reflect the related physics more completely. This is indeed included in our research plan. On the other hand, predictions of GEO electrons by PreMevE are at least as good as other existing prediction models for GEO electrons, as being already demonstrated in our previous work in 2019." ®