Participants of the NASA Frontier Development Laboratory program on August 17 presented projects on the use of machine learning in space. In particular, the teams showed artificial intelligence systems for determining the orbits of potentially dangerous comets and improving maps of the lunar surface. IEEE Spectrum talks about it.
Companies like Facebook or Google use machine learning to translate text or recognize people in photographs, but machine learning techniques are used not only in custom products, but also to solve scientific problems. With the help of the Frontier Development Laboratory program, which is being organized for the second year, NASA is exploring the possibilities of artificial intelligence algorithms for space exploration. Each summer, the agency brings together small groups of researchers to tackle important space research problems.
In total, the teams are working on five projects - protecting the planet from long-period comets, identifying lunar craters, creating three-dimensional models of near-Earth asteroids, studying the effect of the heliosphere and space weather on the Earth's atmosphere and magnetosphere, and determining the causes of solar flares and coronal mass ejections. At the Wrap-Up conference in Santa Clara, which took place last Thursday, scientists presented the first results.
IEEE Spectrum spoke about the results of the work of the two teams. The first team of researchers used data from the Cameras for Allsky Meteor Surveillance (CAMS) survey to predict from meteor showers when the next long-period comet will fly near the Earth. As part of CAMS, sixty video cameras installed at three stations watch the sky looking for faint meteors. They find meteor showers and try to correlate them with recently discovered comets that may have left these debris. A team of scientists from the Frontier Development Laboratory has developed a neural network that distinguishes fast-moving meteors from clouds, fireflies and airplanes (usually done by hand), and then groups the images in time. Thus, the algorithm finds previously unknown meteor showers.
In 90 percent of cases, the predictions of the neural network, which was tested for two months, coincided with the classification of objects by humans. In a pilot project, the team analyzed about a million meteors. However, some experts were skeptical about the project: in particular, they demanded proof that meteor showers are not noise in the data, and also that they are the remnants of comets, and not asteroids or other sources. One of the creators of the project, Marcelo de Cicco from the Brazilian National Institute of Metrology, agreed that the neural network still needs to be improved.
The authors of the second project worked with data from the Lunar Reconnaisance Orbiter (LRO) interplanetary station to create a more detailed map of the lunar surface. Scientists first used information from the Lunar Orbiter Laser Altimeter (LOLA) to create a digital elevation map of the satellite. However, it had one drawback - it contained artifacts. Each time the LRO orbits the Moon, it deviates slightly from its ideal orbit. Because of this, measurements are inaccurate and rocks and cracks appear where they are not.
To solve this problem, the researchers matched the map with images from the Narrow Angle Camera (NAC), which records sunlight reflected from the moon's surface. Using a machine learning algorithm, the team weeded out the artifacts and made a more accurate map of the Earth's satellite. Scientists have also taught an artificial intelligence system to distinguish craters from shadows and similar objects. The accuracy of the program was 98 percent.
Astronomers have increasingly used neural networks in their work in recent years. For example, computer algorithms are already helping scientists determine the composition of the atmospheres of exoplanets and track the movement of stars in the galaxy.
Promotional video:
Christina Ulasovich