CosmicAI Institute Tackles Universe’s Deepest Mysteries

Astronomy offers an exciting frontier for the advancement of large language models and artificial intelligence. These powerful tools are poised to revolutionize the field, helping sift through vast quantities of existing and new data streaming from powerful current and next-generation telescopes. With the cosmos still shrouded in countless mysteries, these tools will help astronomers answer the questions: What is our place in the universe? How did the universe evolve, and what fate awaits it?

The launch of the NSF-Simons AI Institute for Cosmic Origins (CosmicAI), in October 2024 at the Oden Institute for Computational Engineering and Sciences, marked a natural evolution in Stella Offner’s career. As director of CosmicAI and a professor of astronomy at The University of Texas at Austin, Offner now leads efforts to harness artificial intelligence to probe some of the Universe’s most profound mysteries. 

The collaborative team, which brings together seven research entities, aims to improve efficiency, explainability, and accessibility of AI and astronomy methods, with goals of building an astronomy AI “co-pilot” and serving as a nexus for collaboration including education initiatives.

“Our goals are to inspire people, gain a better understanding of the universe, develop innovative open-source AI tools for astronomy, and create educational opportunities,” Offner said. She also co-directs the Center for Scientific Machine Learning, which is also housed at UT's Oden Institute.

Focusing on four fundamental AI themes — trustworthiness, efficiency, interpretability, and robustness — the CosmicAI Institute consists of partnerships across the country: UT Austin, the Texas Advanced Computing Center (TACC), the University of Virginia, the University of Utah, and the University of California at Los Angeles, along with two nationally funded astronomy centers: the National Radio Astronomy Observatory and NSF NOIRLab (formerly the National Optical-Infrared Astronomy Research Laboratory).

On the brink of a data revolution, astronomy is entering its own big data era. With the arrival of next-generation radio telescopes, scientists will be inundated with colossal datasets — each one a terabyte-size data cube — packed with nearly a trillion data points, leaving traditional analysis methods struggling to keep pace. For perspective, according to Offner, the amount of data coming from the Vera C. Rubin Observatory every night is equivalent to watching a streaming service non-stop for two years.

Using AI, researchers hope to unlock this data faster, with smarter ways of extracting insights from this cosmic flood of information and explore ways to combine observational data with computer simulations to understand fundamental parameters of the universe and how to accelerate the modeling of astrophysical systems.

“Figuring out how to use techniques to improve data analysis with smaller resources and to accelerate the speed with which those new technologies can be adopted is one of our thrusts,” said Niall Gaffney, who is the computing director for CosmicAI and the Data and AI Directorate for TACC.

The CosmicAI team is leveraging systems at TACC — Vista, Frontera, and Stampede3 — to create benchmarks and training data for AI models. Having both the expertise of TACC personnel embedded into the working groups and access to the powerful systems is critical to their mission.

Another key initiative in development by the CosmicAI team is an astronomy assistant, or “co-pilot,”  to more easily retrieve and interact with astronomy data, essentially an AI agent like ChatGPT. “Astronomy data represents a small percentage of the overall internet content. What we are building requires specific and high-level expertise in astronomy,” Offner stated.

The early stages of creating an astronomy assistant include establishing benchmarks to test existing models and evaluate performance, which will inform next directions. “This is an evolving process since the LLM generative AI landscape is changing rapidly,” added Offner. 

AI and machine learning models are often opaque — the black boxes — and creating a component of explainability in the case of scientific modeling, along with the ability to propagate uncertainties and make quantifiable predictions, is key. “When the large language model generates an answer, we want to know its reasoning and how it arrives at that answer,” said Offner, adding, “it’s not enough to give an answer — we must also give the how and why.”

Building community trust in AI also means education, accessibility, and training, all of which are baked into CosmicAI’s DNA. “When you don’t understand where a given result comes from and how it was generated, it’s very difficult to trust it,” said Offner.

In addition to annual conferences and boot camps, the institute has established a hybrid seminar series and launched an online graduate certificate program in AI and machine learning with a badge in astrophysics. Given the rapid pace of advancements in AI, expanding educational opportunities is essential to ensure more people can participate in and benefit from these breakthroughs.

“Our goal is to create accessible, interpretable tools and make it easier for researchers and students to understand and engage with AI methods,” Offner said.

Explaining the unexplainable is an intrinsic value of astronomy, and according to Gaffney, provides ideas for where to look, and learn what’s next, and how it applies to astronomy and other sciences. Everyday applications, such as residual rare earth minerals from old asteroid craters, come directly from the cosmos. 

“Your cellphone depends on the fact that asteroids hit the Earth millions of years ago,” Gaffney said. “These craters are a rich source for rare minerals that are used to build things we use in our everyday life such as lithium-ion batteries.”

“Combining simulations of the universe with AI methods allows us to understand the past, present, and future of the cosmos,” Offner concluded. “We are addressing the fundamental mysteries, including the nature of dark matter and the origin of complex molecules needed for life.” 

Q&A with Stella Offner

by Jane Bernhard, CosmicAI Communications Director

Tell us about you. What inspired you to lead CosmicAI?

I majored in physics and math and earned my Ph.D. in physics from UC Berkeley. I developed an interest in high performance computing — and became a computational astrophysicist. Ten years ago, my research group began using numerical simulations created through high performance computing as training sets for machine learning and AI methods. When the NSF-Simons AI Institute program opportunity came up, I was excited to apply.

What kind of research is currently underway at CosmicAI?

One exciting research area is happening in our Explorable Universe group, which works to make LLMs for astronomy research more trustworthy and accurate. We developed an astronomy data visualization benchmark to test existing LLMs on tasks like coding. It is a close collaboration between computer scientists and astronomers.

Why is it important to invest in this work?

For the astronomy community, there are AI benefits to accelerating research and doing data analysis. There’s also a synergy between astronomy research and computer science.

Astronomy is an excellent sandbox to develop AI techniques in a safe and open way. We have an ecosystem with open-source code, literature, and nonproprietary data. It’s not related to health or human society. Anything developed here doesn’t have life-or-death implications, which allows us to experiment and develop new techniques in a safe way.