It isn’t just people who are wrestling with the big questions about the cosmos.
At NASA’s Frontier Development Lab (FDL), researchers are using machine learning to explore whether life could exist on other planets, how to defend Earth from asteroids, and how to spot pristine meteorites on our planet’s surface.
The FDL is an applied AI research accelerator hosted by the SETI Institute in partnership with NASA Ames Research Center. The lab focuses on how AI can tackle some of the hardest problems in space science, and brings together researchers from NASA, the European Space Agency (ESA) and academia with those from Google, IBM, Intel, Lockheed Martin, Nvidia and various other companies.
“Artificial intelligence is really invaluable right across the spectrum of space problems,” said James Parr, director of the FDL, adding that the lab was focused on research both into space and how space-based technologies could solve problems on earth.
SEE: Photos: 60 years of NASA’s technological accomplishments
He evoked the famous Earthrise photo taken as the Apollo 8 spacecraft orbited the moon, showing the earth as a tiny blue bauble emerging into the vast black canvas of space.
“That’s the thing about space work, you start by looking up but you end up looking back. You realise our planet isn’t very big and we have a vast number of problems.
“We have a fantastic new technology with AI, I would invite you to start thinking about how to use your ingenuity,” he told the audience of machine-learning engineers and AI researchers at the RE•WORK Deep Learning Summit in London.
Here’s five ways the FDL is using AI technologies to explore the cosmos.
1. To make informed guesses about alien life
The FDL has been using neural networks to explore what types of alien life could exist on exoplanets, planets orbiting stars outside of our solar system.
Even though these planets are light years away, they can be spotted by space telescopes observing periodic dips in light as these planets pass in front of their parent star, providing clues about each planet’s density, mass and distance from their sun.
The FDL used these spectral clues to train autoencoders and generative adversarial networks (GANs), types of neural network that can generate plausible data. Using these trained networks, the FDL was able to generate 3.5 million possible candidates for alien metabolisms, the chemical reactions that sustain life.
Parr says it’s worth exploring how life on alien planets could differ, pointing out that life is “not just the way it’s evolved on Earth, there’s different possibilities”.
The spectra of light passing through exoplanet atmospheres can also provide hints about the atmosphere’s chemistry and the planet’s climate, and Parr said the lab expected to be able to make even more detailed projections based on data from the space-based Gaia and James Webb telescopes.
2. Detecting exoplanets
While humans have detected more than 3,000 exoplanets, Parr said NASA’s recently-launched Transiting Exoplanet Survey Satellite (TESS) should help us identify far more planets than had previously been possible using data gathered by the Kepler space telescope.
“Kepler was really looking at just a postage stamp and now TESS is going to look at 80 – 85% of the sky. It’s a huge data challenge,” he said, adding that much of the data analysis is still manual.
“One of the projects we did this year was to take that workflow and replace that with an AI workflow,” he said.
He said the FDL team had used data gathered by Kepler to prove the technique worked, and hoped to analyze TESS data in the new year.
3. Helping protect the Earth from asteroids
Before we can protect the planet from the Near-Earth Objects (NEOs) hurtling through space, we need to know what they look like.
But modelling the shape of asteroids and other NEOs based on radar data can take human experts up to four weeks.
“It’s very useful to understand the shape of an asteroid before it enters Earth’s atmosphere,” said Parr, due to how the shape of an NEO can affect its aerodynamics.
“Understanding its centre of mass and tumble is actually a crucial thing in determining how to move it if we needed to.”
By feeding this sparse radar data into trained GANs, the team were able to model NEOs within hours.
“NASA has a massive backlog of shape models, to do so this is proving a great application of technology.”
4. Helping recover meteorites
Finding pristine meteorites after they land on earth’s surface is a race against time before the water, oxygen and chlorine in the Earth’s atmosphere take their toll.
FDL researchers used a homemade dataset of about 35,000 meteorites to train a machine-learning model to spot meteorite samples from above and distinguish them from terrestrial rocks.
Once installed on camera-equipped, meteorite-hunting drone the team found the machine-learning model worked extremely effectively at scouring a debris field in the vicinity of where a meteorite was known to have fallen.
“We’re visited by asteroids all the time, meterorites land on the ground and if we can get them soon enough we can get a pretty pristine sample before they oxidize,” said Parr.
“The drone discovered 16,000 candidate meteorites but what was exciting about this was that out of the 16,000 it determined the actual meteorite. It shows this technology is phenomenally powerful.”
5. Mapping lunar craters that may contain water
Mapping deep craters at the Moon’s poles can help identify which craters may contain frozen water, but it is also a “huge, time-consuming task”, according to Parr.
At the existing rate, manually mapping craters at the poles could take more than 2,000 years he said.
To speed up the process, the FDL and Intel created a game where players would help label images of lunar craters. This dataset was then used to train a convolutional neural network (CNN), a type of network that excels at image recognition, to spot craters at the poles.
Compared to human experts, the trained machine-learning model was 100x faster and had a 98.4% accuracy rate.
“Once we have that inference model developed we get some fantastic results,” according to Parr, who said the trained model essentially removed the need for humans to do the analysis.