Unveiling the Universe’s Origins and Current State by Modeling Its Many Faces

Just as cities, communities and people have histories, so too does the universe. The galaxies and stars we see today took shape over billions of years of cosmic evolution.

Using data science and machine learning, Adrian Bayer of the Flatiron Institute’s Center for Computational Astrophysics (CCA) reconstructs the past, interprets the present and forecasts the future of the cosmos.

Bayer started as a research fellow this year at the CCA and is the co-leader of the Simons Observatory’s Sky Modeling working group. He holds a doctorate in physics from the University of California, Berkeley, a master’s in mathematics from the University of Cambridge and a bachelor’s in physics from Imperial College London. Before joining the CCA full-time, he worked as a postdoctoral researcher at Princeton University.

We recently discussed Bayer’s work and the value of learning from as many researchers as possible. This conversation has been edited for clarity.

What inspired your interest in cosmology?

In high school, I was very interested in how things work, particularly the inner workings of the human body and the structure of the universe. The body is at a very different scale than the universe, of course, but for me, understanding the dynamics of both was the key to understanding our place in the world.

Eventually, I had to choose between biology and physics. I arrived at physics, as it offered more opportunity to use mathematical and computational skills to solve problems, which I knew I would enjoy and find rewarding. I landed on cosmology after doing research in particle physics and quantum gravity because it’s a nice blend of theory and observation; as a cosmologist, I can apply the models and methods I develop to observations that are happening as we speak.

What is the focus of your current work?

I use data science and machine learning techniques to understand how galaxies are distributed and formed throughout the universe. I also study the cosmic microwave background — a relic light from the early universe, about 380,000 years after the Big Bang, now observed as microwave radiation that travels through space.

As co-leader of the Sky Modeling group for the Simons Observatory, I work on modeling how the microwave background interacts with matter and galaxies on its journey to us. Another key component of my work is using these observations to rewind the evolution of the universe to find what it looked like moments after the Big Bang.

This work involves running state-of-the-art simulations on supercomputers and analyzing their output. Preparing the simulations involves inputting all variables we are interested in — from the amounts of dark matter and dark energy, to the mass of tiny neutrino particles, to rules that govern gas, stars, and black holes.

Once the simulations are run, we analyze them as we do the real sky: We turn them into mock pictures and catalogs and line them up with telescope maps to see where the simulations match with the real universe, in turn teaching us how the universe fundamentally works.

Machine learning helps in two ways: It learns subtle patterns — tiny temperature shifts in the microwave background, the “shadows” of galaxy clusters, and the web-like arrangement of galaxies — that are easy to miss by eye, and it can rapidly test millions of possible “recipes” for the universe to find the ones that best match reality. With this approach we answer big questions: How does dark matter sculpt the cosmic web? How fast is the universe expanding, and what role do ghostlike particles called neutrinos play? Where is the hard-to-see normal matter between galaxies? Just as important, these tools tell us how confident we are in each answer, so we can design better observations and steadily sharpen our picture of the cosmos

How does the Simons Foundation support these efforts?

The Simons Foundation makes this work possible in three big ways. First, it provides the computing and data infrastructure we need to grow “model universes” and compare them to real telescope maps at scale. Second, dedicated research engineers help turn ideas into reliable, efficient code and keep our pipelines reproducible — so results can be trusted and shared. Third, the culture is deeply collaborative: Working groups bring together cosmologists, data scientists, and engineers, and that mix routinely sparks new methods, better tests, and faster progress.

It’s not just about computing power; it’s about people and openness. We use shared machine learning tools, common data standards and open-source software so that what we build becomes a resource for the broader community. The foundation’s support lets us take on ambitious, long-term questions about the universe with the confidence that both the expertise and the infrastructure are there to see them through. More broadly, I find that it’s easy to strike up conversations at Simons about what I am doing and also what others are working on. Sometimes, those casual chats lead to new projects.

Describe one of the projects that began as a casual chat.

One recent project was to develop an agentic AI tool that can help scientists sift through literature, generate ideas, write code and prepare scientific documents. After randomly chatting with colleagues at Simons about their work, I became interested in agentic AI and what it might be able to do, so I decided to develop a refined idea-generation process for such an agentic AI system. We call this the “idea maker” meeting the “idea hater” — where ideas are improved iteratively through conversation between different agents, just like humans would in the real-world scientific process.

Through this project my collaborators and I ultimately developed an AI that we call Denario. We have now applied this pipeline to various topics of research, from astrophysics to biology, at the centers of the Flatiron Institute and beyond.

Another great example of collaboration comes from our astrophysics and mathematics centers, where we teamed up with machine learning researchers to build CosmoBench — a set of apples-to-apples tests that compare machine learning and physics-based methods on key cosmology problems. CosmoBench has been accepted for presentation at NeurIPS, a leading international AI conference.

AI has many uses in science beyond this. The machine learning that I use for my cosmological simulations is a form of AI, where we use it to speed up modeling, learn how to undo the evolution of the universe and obtain more precise scientific results.

Finally, please expand on the importance of collaboration in your work.

Collaboration is how we turn many partial views and skill sets into one reliable picture of the universe. In practice, that means physicists, data scientists and software engineers designing models together, stress-testing each other’s work and assumptions, and sharing tools so our results can be checked and reused.

It also means being open to other ways of working and seeing the world. I’ve been fortunate to have the chance to travel to do research throughout my career — for example, as an undergraduate in London, I had a summer research visit at the Massachusetts Institute of Technology (MIT), and as a Ph.D. student in the U.S., I did a research visit to Tokyo. Those visiting scholar opportunities gave me a chance to observe other scientific cultures and ways of thinking about the problems I work on day by day.

The Flatiron Institute acts as a hub, attracting many leading national and international researchers, so new ideas and conversations often come to you. I’d encourage anyone in any scientific field to look for chances to learn from as many people as possible; that is one the surest paths to being a top scientist.