Updates to our Terms of Use

We are updating our Terms of Use. Please carefully review the updated Terms before proceeding to our website.

Monday, June 24, 2024 | Back issues
Courthouse News Service Courthouse News Service

Machine Learning Helps Penetrate Galaxy’s Mysteries

(CN) - A machine learning simulation process developed by researchers from the University of Illinois could make galaxy modeling more efficient and offer astronomers an opportunity to derive new insights into how the universe evolves.

Galaxy simulation ordinarily requires countless computing hours, making the process costly and inefficient. To develop accurate models that derive new findings, numerous simulations that factor in known and unknown variables must be run.

"When we make a cosmological measurement, we have some implicit uncertainty since we can't rerun the experiment," University of Illinois astronomy professor and process co-author Robert Brunner said. "One way to characterize this error is by simulating multiple universes and seeing how the measurements change. But this is a computationally expensive prospect, and takes a long time. I want to speed this up."

Brunner, along with undergraduate student Harshil Kamdar and scientist Matthew Turk, developed machine learning, which runs algorithms that can recognize relationships between complicated sets of data. This allows a researcher to input specific properties or factors in order to create expeditious models.

"Basically we train an algorithm on some fraction of the data and predict on unseen data. This is similar to how Amazon 'knows' how to recommend certain items to you, or how loan applications can quickly be screened," Brunner said in an email.

Researchers primarily use two separate models to simulate galaxies.

The first method incorporates hydrodynamical N-body simulations, which models the interaction between normal matter and dark matter under the gravitational stress that leads to the creation of galaxies and the formulation of stars.

The other strategy researchers use is N-body simulation paired with a semianalytical analysis, which details how dark matter reacts to collapses under gravity and describes how galaxies form. Such projections require super computers and thousands of hours.

Machine learning algorithms can assist in bypassing many of the hurdles that have limited astronomers using these N-body simulations, which could lead to an extensive repository of projections.

"The goal in the end is to create mock galaxy catalogs," Brunner said. "Right now we are still showing it is possible, but more work remains. We want mock catalogs to quantify the errors on cosmological quantities from measurements we make of the 'real' galaxy distribution."

While the system is largely accurate, some data points and physical properties are not fully known at this point which can partially limit the projections.

"We don't have sufficient input data to learn how to model all galaxy properties. There is complicated physics at play, and our algorithms currently have insufficient data on hand to fully model all of that," Brunner said.

Machine learning could provide an opportunity for astrophysicists to gather more data and conduct enhanced research and produce diagnostic tools for cosmological simulations that require millions of CPU hours to verify that the processes are running properly.

"We are exploring more powerful learning approaches to recover some of the more complicated physical relations that drive galaxy formation and evolution," Brunner said.

Machine learning could be beneficial to other scientific disciplines, as the algorithms can be applied to different sets of data and variables.

"Machine learning is applicable to any domain. You simply need sufficient data with inherent relations that can be modeled," Brunner said.

The researchers will begin to produce a series of mock galaxy catalogs from an existing simulation before deciding how to proceed further.

Turk, who works for the National Center for Supercomputing Analysis, provided scientific oversight for the project, and is a simulation expert. Kamdar organized several key components of the experiment.

The center declined a request for interview.

"We do not feel your publication fits our strategic goals at this time," said Barbara Jewett, the center's managing editor of public affairs.

Follow @SeanDuffyCNS
Categories / Uncategorized

Subscribe to Closing Arguments

Sign up for new weekly newsletter Closing Arguments to get the latest about ongoing trials, major litigation and hot cases and rulings in courthouses around the U.S. and the world.