It's Extraordinarily Hard to Weigh Galaxy Clusters. AI Just Made It a Breeze.
Galaxy clusters are the largest structures in the known universe, sometimes containing more than 1,000 galaxies within them.
Understanding the masses of these clusters gives astronomers vital information regarding the formation of the universe.
A team of scientists leveraged an AI tool called “symbolic regression” to more accurately “weigh” these massive objects.
Part of understanding the universe involves knowing how much stuff is in it. Sometimes, this means figuring out the weight of galaxy clusters, the largest celestial objects in existence. Galaxy clusters can contain upward of 1,000 galaxies, and understanding their total mass aids in our knowledge of how the universe formed.
As you’d expect, “weighing” these objects is no easy feat. It requires advanced mathematics to account for all the plasma, hot gas, and dark matter swirling inside. For decades, scientists have used equations based on the idea that electrons interact with photons differently under different amounts of pressure—which is caused by gravity. By analyzing the new photons produced by those photon-electron interactions, scientists can estimate the amount of gravitational force produced by a cluster, and from that, estimate a cluster’s mass.
However, this method isn’t perfect, because a photon’s properties can change based on the galaxy cluster they’re in, not just the pressure they’re under. So, astrophysicists at the Institute for Advanced Study and the Flatiron Institute have leveraged AI to make a better equation for weighing these impressive celestial structures. Surprisingly, the AI added only one simple term to the existing equation, and in doing so made the measurement much more accurate. The team published their results in the Proceedings of the National Academy of Sciences (PNAS) last week.
“It’s such a simple thing; that’s the beauty of this,” says Flatiron Institute research scientist and co-author Francisco Villaescusa-Navarro in a statement. “Even though it’s so simple, nobody before found this term. People have been working on this for decades, and still they were not able to find this.”
Many researchers employ a commonly used type of artificial intelligence called a “deep neural network” to help with their calculations and data analysis. However, As the researchers describe, neural networks can be a bit of a “black box”—the machines can come up with amazing results, but it’s often difficult to figure out how they arrive at their conclusions.
So, instead of using a neural network to make their calculations, the researchers used an AI tool known as “symbolic regression.” This technique uses existing data sets to produce results that are easily understandable and can, crucially, be reverse engineered to figure out how the AI came up with a solution.
The scientists fed their symbolic regression program highly advanced simulations of the universe populated with galaxy clusters and tasked it with identifying variables that would improve the accuracy of mass estimates. After crunching numbers, the AI spit out only a single new term for an existing equation.
As the researchers reverse-engineered the result, they discovered what the AI had noticed—in trying to estimate the overall mass of everything in a cluster, too much importance was being placed on regions where mass estimates were the least reliable due to gas concentrations. This included regions like the centers of galaxies, which are often home to supermassive black holes.
In the new equations, the AI downplayed the how much those regions matter within the overall calculation. It doesn’t remove these regions from the data set, but it does effectively screen out their errors, making a much more accurate mass estimate.
The paper’s co-authors think this is only the beginning of exploring the utility of AI tools like symbolic regression in astronomy, and believe that AI can be a major help in exploring everything from the complexities of small exoplanets to the grand majesty of galaxy clusters.
You Might Also Like
