When cosmology makes headlines, we often see flashy images of cosmic maps and supernovae. But in reality, scientists must spend months or years sifting through hundreds or thousands of calculations and simulations. To ease this burden, some scientists are turning to AI, but new research shows the benefits and drawbacks are fairly nuanced.
In a study published earlier this month in the Journal of Cosmology and Astroparticle Physics, cosmologists trained an AI neural network on simulations of the standard model of cosmology, ΛCDM. The team then tested whether this pre-training helped or hurt the AI’s subsequent exploration of other open questions in cosmology and astrophysics. Although AI showed some promise, it introduced biases that ultimately hurt the discovery of new physics.
The study is “a great example of how AI, when used in a structured way, can help advance scientific progress faster,” study co-author Adrian E. Bayer, a cosmologist at the Flatiron Institute and Princeton University, told Gizmodo. “At the same time, this study reminds us that acceleration and understanding must go hand in hand.”
expensive truth
Cosmological breakthroughs tend to be expensive and time-consuming. As Dark Energy Spectroscope Instrument (DESI) co-spokesman Will Percival told Gizmodo in April, preparing a dataset for scientific analysis requires creating mock universes and galaxies and running simulations as a sanity check. These processes are essential for drawing significant conclusions from sophisticated observations.
But simulations of the model beyond the Standard Model (extensions that include things like massive neutrinos, evolving dark energy, and modified gravity) are also very expensive, Bayer told Gizmodo. At the same time, testing these alternative scenarios, regardless of whether they are ultimately correct, is important to advancing our understanding of the universe. This practical motivation led Bayer to search for “an efficient way to learn for every scenario without requiring a large new simulation suite.”
Bumpy transfer?
In the experiment, the team used a machine learning strategy called transfer learning. In this approach, a model first learns from one task or dataset (a simulation of the Standard Model) and then applies this knowledge to learn extended versions of the Standard Model that include related tasks and promising ideas for new physics.
According to Bayer, AI performed significantly better at understanding the standard model based on fewer, lower-cost simulations. But when the new physics “combined with directions we had already learned, we started to struggle.” [the standard model] “This phenomenon, called negative transfer, occurred when the AI became biased and unable to distinguish between two different physical effects that produced similar patterns in the data. So instead of finding something inherently new, the AI relied on what it had already learned and missed potential clues that suggested physics beyond the standard model.”
“This negative transfer result is interesting because it shows that the model is not failing randomly,” Bayer added. “Understanding when transfer learning is useful and when it strengthens degeneracy is critical to ensuring the use of AI in future cosmological analyses.”
AI and cosmology
For Bayer, the latest findings confirm a not-so-new concept that AI can help, but human experts will need to carefully follow its calculations to understand and pursue the relevant questions.
“Transfer learning gives AI a strong head start, allowing it to test more ideas than practical ideas about the universe,” he said. “But if a model carries over knowledge from one setting to another, we need to understand what that model carried over, when that knowledge is useful and when it can be misleading.”
Next, Beyer and his colleagues plan to conduct similar experiments in settings that “more closely resemble real survey data,” including “uncertainties in galaxy formation, survey masks, and noise.” Additionally, the team would like to investigate which cosmological lines of inquiry would benefit most from transfer learning.
