Artificial intelligence and the structure of the universe
We’ve all seen artificial intelligence in sci-fi movies and television. It seems the “machines” are getting smarter every day! But how can we harness artificial intelligence to help us learn about our universe, and possibly even extraterrestrial life?
Machine learning is a subfield of artificial intelligence that describes how a computer program can take information, identify patterns and learn to identify the important aspects lurking in the data. One example is a neural network (“neural” because its learning is patterned after the neurons in the human brain). If we give a neural network millions of images of animals and an answer key of which animal is which, the neural network will soon learn to identify animals in pictures. Now give it millions of different pictures of these same animals and it will identify them all for you in a second. Imagine if we unleash that lightning-fast precision learning on the study of the universe.
So how do we teach a machine if we don’t have millions of animal pictures on hand? Well, how might we teach a child? If you didn’t have a picture book of animals to teach your child, you might draw your own. This is one of the most commonly used solutions in machine learning. If the training data don’t exist, we simulate them. In order to properly teach a machine about a topic, we’ll need the most detailed, accurate simulations possible which means that we need to pool our complete knowledge on the subject. Otherwise, it can be easily tricked by “lookalikes,” such as in the famous “dog or food?” internet meme. Imagine blueberry muffins masquerading as chihuahuas!
Now consider space: full of stuff like stars and galaxies. Galaxies, including the Milky Way, reside in large halos of gas that astronomers call the “circumgalactic medium.” Beyond the circumgalactic medium, filaments of gas stretch between galaxies. This gas comprises what we call the “intergalactic medium.” Together, circumgalactic and intergalactic gas form a large overall structure that astronomers call “the cosmic web.” The cosmic web, massive filaments of galaxies separated by giant voids, is the structure of the universe.
A powerful technique to reveal the structure of the cosmic web is to observe the light from very distant and very bright objects. As this light travels billions of light years through the universe on the way to our telescopes, it passes through the gas that makes up the cosmic web; the gas absorbs some of the light, leaving a distinct signature that we can measure. The signature contains information about the chemical composition of the gas in the cosmic web, how much gas there is, its temperature, and more. By analyzing many of these systems, we can start to piece together information about the structure of the universe, and how it has changed throughout its lifetime.
As powerful as this technique is, it is equally complex and time-consuming. In fact, a trained expert might analyze only one or two systems per week. For reference, the Astronomy Department at NMSU has around 3,500 systems on file, which means that it will take 35 to 70 years to analyze all of this data. Furthermore, our data archives will only grow as the next generation of telescopes come online. How can we possibly handle all of this data? Enter machine learning.
Using machine learning, it may be possible to leapfrog the system analysis entirely. Instead of feeding a machine pictures of animals to identify, we can feed it absorption systems and receive our scientific answers directly from the machine. This work is still in the beginning stages of development and will take many years to pursue, but the potential of applying machine learning to quasar absorption line spectroscopy is incredible. With our current models, we can train a machine on nearly one million simulated systems in just over a day, at which point the machine can analyze 100,000 systems in just a few hours!
As our telescope technology improves and we collect more data in greater detail, astronomers face the daunting task of finding new ways to analyze these data en masse. Machine learning holds significant promise in this effort and is justifiably being explored in astronomy as a new tool available in our endeavor to understand the cosmos.
Bryson Stemock is a PhD student in astronomy at New Mexico State University. He can be reached at bstemock@nmsu.edu.
More Star News: