Light years away, dense stars are blazing into fiery stages of life as red giants.
It’s challenging to determine the exact phase of these aging stars in the last stages of stellar evolution.
A team of scientists from Australia and Denmark trained a deep learning system to predict the phases of existing red giants identified by NASA’s Kepler mission. The GPU-powered system greatly sped up their work, and delivered remarkable accuracy.
AI Accelerates Asteroseismology
Red giants have two phases of life that look very much alike on the outside. A younger red giant burns hydrogen outside its helium core, while an older red giant has started to burn the helium in its core.
To figure out the difference is where asteroseismology, a science that studies the internal structure of stars by interpreting their frequency spectra, comes in, according to Marc Hon, a Ph.D. student at the University of New South Wales and study co-author.
“The composition of the star can be determined by the frequencies at which a star vibrates,” said Hon. “We measure these frequencies by observing very tiny light fluctuations coming from the star.”
By visualizing these frequencies on a frequency plot known as a power spectrum, it can be seen that the power spectrum of the core-helium burning red giant displays a messier pattern than the hydrogen-shell burning red giant. It could take a few seconds, depending on the pattern presented, for a human eye to deduce the phase of one star behind the spectrum.
That may not sound like a long time, until one remembers that there are literally hundreds of billions of stars in the Milky Way galaxy alone. That’s why the researchers turned to a deep learning algorithm.
“If I were to do it manually, it would take me at least 5-10 seconds just to identify one star,” Hon said. “In contrast, the deep learning algorithm could predict thousands of stars in a matter of seconds.”
AI the Expert Eye
Researchers decided to convert the spectrum data into an image-like representation that a deep learning system could recognize. This resulted in greater data analysis, faster feedback and a 99 percent accuracy rate.
As a result, the AI classified 7,600 change red stars, about 5,400 of which had not previously been analyzed. It’s also corrected 500 manual star classifications.
The system was built using a cuDNN (CUDA deep neural network) library and trained on a workstation powered by NVIDIA Quadro graphics.
AI’s Future in the Field
Since the frequencies on the power spectrum reveal more than just the phases of red giants, Hon said the AI could be trained to interpret more information, such as whether the star is vibrating in the first place.
By obtaining the phases of these aging stars, their study can be useful for researchers mapping the locations of red giants to trace the structure and evolution of the Milky Way.
“We can use deep learning to look at other features on the power spectrum, so there is potential for AI to tell us a lot of other things about the stars,” Hon said.
For now, their deep learning algorithm will continue to become more precise as the data set grows, allowing them to circumvent tedious traditional practices in their field.