How Arya Farahi Uses Simulated Universes to Advance Dark Matter Research

For nearly a century, one of physics' most stubborn mysteries has concerned the properties of dark matter. Invisible yet noticeable through its gravitational effects, dark matter makes up, theoretically at least, about 85% of the total matter in the universe. What is it, and how exactly does it work? 

Arya Farahi, an assistant professor in UT Austin's Department of Statistics and Data Science (SDS), spends his days building computational tools that might finally help answer those questions — not by discovering a new particle in a lab, but by teaching algorithms to compare thousands of simulated universes against the one we actually live in and observe.

"There are a lot of theoretical models proposed, but we have not been able to exactly constrain the properties of dark matter and say which of these models can explain our observational data better," Farahi says. His hope is that, through AI-assisted statistical techniques, he can help cut through the messy noise of observed data of distant galaxies and shed some light – so to speak – on this most obscure area of astrophysics.

Why the Universe Needs Data Scientists

Ask Farahi what role statisticians play in modern cosmology, and he'll point you to the data avalanche. Cosmological surveys and simulations now generate volumes of information that scientists struggle to use effectively. "The level of data we are producing, both through numerical simulation and observational data, is exceeding the capacity of existing traditional and statistical models to analyze, extract patterns and make sense of," he says. 

Half of that data comes from massive "boxes"; simulated, digital universes built under different physical assumptions, each one including a large population of simulated galaxies that can be checked against what telescopes actually see. "For each simulation, we are generating tens of terabytes of data,” he says. “And imagine now that we have tens of thousands of simulations."

Complicating matters further, real observational data is never clean. Telescopes scan the sky unevenly; instruments introduce noise. Part of Farahi’s research focuses on building inference methods that can cut through that noise, allowing other researchers to meaningfully compare simulations of our universe with observations from real surveys.

The DREAMS Project: Simulating Alternate Universes

The centerpiece of Farahi's current research is his work with the DREAMS Project (DaRk mattEr and Astrophysics with Machine learning and Simulations), a multi-institution effort — including researchers at UT, MIT, Princeton, CERN and elsewhere — that builds cosmological simulations designed specifically to probe the nature of dark matter.

Most cosmological simulations assume a single dominant model for dark matter, known as Lambda-CDM. DREAMS explores a wider range of possibilities. "These alternative dark matter models are typically not simulated, so we don't have a good predictive model for them,” Farahi says. 

DREAMS fills that gap, generating universes under a range of alternative dark matter assumptions, then analyzing the resulting populations of galaxies. "Our first goal is to identify patterns: What properties of galaxies are super-sensitive to changes in the dark matter properties, and what properties of galaxies are potentially robust to these changes," he says

Farahi is careful not to overpromise. "This is a 100-year-old problem. I'm not sure we are going to solve this problem in five years," he admits. Instead, he frames DREAMS as groundwork: building the infrastructure and statistical approaches the field needs to eventually make progress. "We are trying to prepare the community to answer these hard questions," he says. “Let’s create machinery that can make sense of the dataset."

Teaching AI to Build Universes

Within CosmicAI — a five-year, multi-institution initiative funded by the NSF and the Simons Foundation, where he leads one of four "grand challenge" research areas — Farahi’s primary contribution is developing the AI techniques that help researchers better use simulation data.

But what if thousands of simulated universes aren't enough to solve the underlying structure of the universe? Increasingly, Farahi has been working with generative AI to scale up simulations. "We are training generative AI that can learn from the patterns of existing simulations and regenerate new universes on a scale of a few seconds," he explains. "As soon as you want to change the universe parameter, you can, in a few seconds, generate a new simulation."

This ability to rapidly regenerate simulated universes is also useful as a way of comparing observed and theoretical data when no traditional likelihood function applies. Farahi calls this AI-forward approach “simulation-based inference.” 

“You change the parameters behind the simulation until they match the observational data, and you use that engine to quantify the level of uncertainty on the parameters that you want to estimate,” he explains. 

An Interdisciplinary Home in SDS

Farahi sees his cross-disciplinary approach as characteristic of the department where he makes his academic home. "We are trying to build a department that is multidisciplinary, working with the domain scientists," he says. “Often, the domain scientists are the people with the interesting questions, but they don't have the most advanced and appropriate computational tools to analyze their data."

He’s also optimistic about what joining UT's new School of Computing will mean for SDS. "Bringing people from very different computational domains together to think about big problems can accelerate the way we think about computational science," he says.

Whether or not the nature of dark matter is finally pinned down in the years ahead, Farahi's work is reshaping how major cosmological mysteries are solved, turning an old physics question into a frontier for modern data science.

Click here to learn more about how researchers at CosmicAI use cutting-edge technologies to advance the frontiers of human knowledge.