How Deep Learning Will Speed Search for Extraterrestrial Life

Those gazing into the night sky have speculated about life beyond Earth since Zeus was a boy. Deep learning now holds the promise of zeroing in on an answer.

A deep learning system devised by astronomers at University College London sifts through data from telescopes trained on faraway solar systems to detect planets with the potential to sustain life.

“We want to know which planets are worth further study and which aren’t, and we want to automate that,” said Ingo Waldmann, the University College London post-doctoral researcher who leads the development team.

He calls the GPU-accelerated deep learning program RobERt, short for Robotic Exoplanet Recognition. Exoplanets are those beyond our solar system. Deep learning, a type of artificial intelligence, is a way of training computers to perform tasks like image or speech recognition with near-human accuracy.

How Deep Learning Detects Signs of Life in Space

Using data collected by telescopes, RobERt examines light passing through the atmospheres of exoplanets for signs of methane, carbon dioxide and other gases linked to biological activity.

Why light? Different types of molecules absorb and emit light at specific wavelengths, so each has its own “fingerprint” in a light spectrum. That fingerprint, or pattern, tells astronomers what types of gases are present.

“We humans are very good at finding these patterns and labeling them from experience, but it’s a really time-consuming job,” Waldmann said. “What usually takes days or weeks takes RobERt mere seconds.”

Waldmann and the team trained their deep neural network on more than 85,000 simulated light wavelengths and five types of exoplanets using CUDA Python with NVIDIA Tesla K80 GPUs and Tesla K40 GPU accelerators. Each light spectrum had the fingerprint of a single type of gas.

kepler186f-search-for-life-beyond-earth
Kepler-186f was the first rocky planet found within a habitable zone. Image courtesy of NASA Ames/JPL-Caltech/T. Pyle.

AI Aids Understanding of Distant Solar Systems . . .

The researchers have used RobERt on data coming from the Hubble Space Telescope. The real challenge, said Waldmann, will come in the next decade with the launch of more powerful telescopes and ambitious space missions. The first of these, the James Webb Space Telescope – an infrared telescope that’s designed to observe the first galaxies formed in the Universe – is set to launch in 2018.

“The amount of data these missions will provide will be breathtaking,” Waldmann said. “RobERt will play an invaluable role in helping us to analyze that data and find out what distant worlds are really like.”

Waldmann’s goal is to have an automated system that takes data  from the telescope and delivers an analysis of the planet’s gases, molecules, climate and chemistry. Already, scientists have learned that our own solar system isn’t the norm in the galaxy, which contains everything from frosty ice worlds to large, scorching planets such as “hot Jupiters.”

. . . As Well as Our Own

With this sort of data, scientists can build a better understanding of how distant solar systems were formed. Being able to look at a statistical sample of different solar systems will shed light on how our solar system came into being, a process that is poorly understood, Waldmann said.

“Before 1995 we only knew of one solar system, ours. Now we know of over 2,500,” Waldmann said. “With so much to look at, we will increasingly need AI systems like RobERt to make sense of it all and put our solar system in the grander context of our galaxy.”

Similar Stories