Key Takeaways:
- AI improves the precision of cosmological parameter measurements.
- This AI method is significantly more cost-effective than traditional approaches.
- The AI analyzes both large and small-scale galaxy distribution data.
- Improved precision helps resolve discrepancies in universe expansion rate measurements.
For almost as long as humans have existed, we have been trying to make sense of the cosmos. What started as philosophical musing has, following the advent of the telescope and the ability to look ever farther into space (and ever earlier in time), become a thriving field of research.
Today, scientists seek to understand the properties governing how our universe behaves. These properties are characterized mathematically as so-called cosmological parameters, which fit into our models of the cosmos. The more precisely these parameters can be measured, the better we are able to differentiate between models, as well as validate — or rule out — long-held theories, including Einstein’s general theory of relativity. Because different models can hold vastly different predictions for both our universe’s earliest moments and eventual fate, that differentiation is vital.
To date, some of the biggest challenges include more tightly constraining parameters such as those that determine the precise amount and nature of dark matter, the source of dark energy and the repulsive force that it exerts, and exactly how neutrinos behave.
These questions are at the forefront of the field of cosmology. However, there is a catch: Probing these cosmic parameters is an expensive affair.
Reducing costs
“There are so many experiments and astronomical surveys built just to measure these six to 10 [major cosmological] parameters and they cost multiple billion dollars,” says Shirley Ho, professor of astrophysics at the Flatiron Institute in New York.
In a paper published last August in Nature Astronomy, Ho and her colleagues leveraged artificial intelligence (AI) to calculate five of the major cosmological parameters governing dark matter, dark energy, and neutrinos to a higher degree of precision than ever before. They did so not only to show AI can make such calculations more precise, but also to show this method is more cost-effective as well. “It’s quite interesting, what they’ve [done] in the paper,” says Emily Hunt, an astronomer at the Max Planck Institute for Astronomy in Germany who was not involved with the research.
Probing cosmological parameters typically requires studying how galaxies are distributed and their individual properties using surveys. Previous approaches used simplified models of the universe and compared them to survey data. This approach mainly allows for a comparison of “blurry,” generalized models to the survey data, ignoring smaller details due to the expense and challenges associated with developing high-resolution models. Trying to look at the universe this way, however, is “like wearing really bad glasses,” says Ho.
Insights from AI
The new study used an AI framework called Simulation-Based Inference of Galaxies, or SimBIG, to extract cosmological information. The scientists first trained the AI model by presenting it with 2,000 simulated universes, each with different cosmological parameters. The researchers then introduced noise to make the artificial universes look more like what we’re able to observe in our own, due to inherent uncertainties in data gathered from galaxy surveys. The noise mimicked natural imperfections introduced by telescopic instruments and the atmosphere (such as bright stars and objects that are so close together that their signals blend).
Over time, the AI model was able to learn to extract hidden features. While previous approaches were only able to look at large-scale distribution of galaxies, the new AI model learned how to leverage small-scale differences in galaxy distribution — for instance, the distance between individual galaxy pairs —- to better estimate the desired cosmological parameters to higher degree of accuracy.
The trained AI model was then presented with more than 100,000 galaxies from the Sloan Digital Sky Survey’s Baryon Oscillation Spectroscopic Survey (BOSS). (This is just a mere fraction of the total BOSS survey — less than 10 percent.) The AI was then able to use both large- and small-scale information in the real data to more precisely constrain the cosmological parameters compared to previous methods. “The advantage of machine learning is that you can just throw these really powerful algorithms at data sort of blindly, and it might extract details that are not something that you would’ve thought of previously,” says Hunt.
The AI-driven approach is critical, since we don’t have another universe, says Ho. “Our best bet is actually to increase the precision as much as we can to squeeze out as much information as we can from the existing universe that we have observed,” she adds.
Future surveys will be able to capture ever-more information about the universe. And AI-powered insights such as the ones offered by this study may be useful in resolving some of the biggest cosmological conundrums today, including the so-called Hubble tension that surrounds the discrepancy in the measured rate at which our present-day universe is expanding. This rate is given by the parameter H0, called the Hubble constant. “If you have a much better constraint [on H0], then you can really nail down whether there is true [Hubble] tension or discrepancy,” says Ho.
Opportunities and challenges ahead
To catalyze further machine learning-based research in astronomy, Ho and co-author Liam Parker (also at the Flatiron Institute) have teamed up with other astronomers to curate a large-scale collection of astronomical data. This effort combines hundreds of millions of publicly available datasets from major astronomical surveys in an effort to “enable the development of large multi-modal models specifically targeted towards scientific applications,” the authors write in an abstract available on the ArXiv preprint server.
While the future of AI’s application in astronomy seems promising, experts urge exercising caution when implementing these tools. “With every breakthrough, certain things kind of race ahead and then other stuff has to catch up,” says Hunt. Understanding of the limitations and uncertainties of machine learning models is critical when applying their results to real-world applications — particularly when attempting to unlock the cosmos.
