Powering Discovery: How Advanced Computing is Transforming Research Across Disciplines

Stepping into the subterranean room that houses the DSI computing cluster, and its many shelves of servers, you’d be forgiven for wondering if you’d just entered a wind tunnel. But that whirring is enabling pioneering work: for researchers across the University of Chicago, this GPU cluster is pushing the boundaries of knowledge across disciplines. From simulating thousands of years of weather systems in order to predict extreme climate events, to forging new ground in understanding human behavior, access to high-performance computing is letting researchers across campus tackle some of the world’s most complex research problems.

But what exactly is a GPU cluster, and why does it matter? It’s a collection of interconnected computers each equipped with specialized processing units (GPUs) that work together as a single system to solve problems. Together, they can process massive datasets, run sophisticated machine learning models, and perform calculations that would take your typical laptop weeks or months to complete.

“In launching the cluster, we hoped researchers would feel empowered to take on bigger and more complex questions than was previously possible,” said Mike Franklin, DSI Faculty Co-Director and Morton D. Hull Distinguished Service Professor of Computer Science, who was part of the team that first proposed the undertaking. “Looking at the outcomes, they demonstrate how computational resources have really transformed what’s achievable on campus.”

While not every research project the cluster contributes to, either directly or indirectly, can be directly identified, authors have called out the cluster’s contribution in over 90 publications. Featured in major artificial intelligence conferences (including NeurIPS, ICML, and ICLR) and high-impact journals (among them Physical Review B, Journal of Computational Physics, and SIAM Journal on Mathematics of Data Science or SIMODS), some of this work is already influencing how we think about AI and research methodology across domains.

Here’s a glimpse into the diversity of research across campus leveraging high-performance computing:

Predicting Extreme Weather in a Changing Climate

In recent years, advances in computing have exponentially increased the possibilities of weather forecasts. Pedram Hassanzadeh, Associate Professor of Geophysical Sciences and Faculty Director of the AI for Climate Initiative, and his team are using the cluster to investigate how machine learning can improve weather forecasting to better understand and predict our changing climate.

“The speed, storage, and staff support of this cluster allow us to bridge traditional atmospheric physics with cutting-edge machine learning approaches: we can train neural network-based atmospheric models on observations and high-resolution simulations, generate many weather realizations with them, and robustly study variability, especially that of extreme events,” said Hassanzadeh. “We can also conduct controlled experiments, for example by training these neural networks many times with carefully designed datasets, to do hypothesis testing. These approaches would have been prohibitively expensive just a few years ago.”

By leveraging these computational advances, Hassanzadeh and his team are developing novel learning strategies to keep pace with climate change to predict and prepare for weather events.

Understanding How the Brain Processes Real-World Information

Historically, study of the brain was limited to relatively brief responses to stimuli, but computational advances have exploded the possibilities.

YC Leong, Assistant Professor of Psychology and Director of the Motivation and Cognition Neuroscience Laboratory at the University of Chicago said, “The DSI computing cluster has been transformative for our research. We are a computational cognitive neuroscience laboratory that studies how the brain processes complex, real-world experiences using full-length movies and TV episodes,” explained Leong. “The cluster allows us to integrate state-of-the-art vision and language models with massive fMRI datasets to predict how the brain responds to rich and naturalistic audiovisual content. No individual lab could maintain the level of storage and computational power needed to run these large-scale analyses. Without the DSI Cluster, tackling questions at this scale, with the latest computational tools, would simply not be feasible.”

By studying the brain under more naturalistic circumstances, his Lab is investigating how people understand social narratives and make predictions about others’ behavior, examining everything from how hostile attribution bias shapes neural responses, to how emotional arousal enhances memory formation during storytelling.

Capturing the Subtleties of Language

Machine learning is also enabling more complex linguistic analysis than ever before. Chenhao Tan, Associate Professor of Computer Science and Data Science and his team are exploring applications across a range of disciplines. As just one example, Tan’s research group developed a framework for analyzing political discourse that makes concrete the uniquely divisive nature of presidential communication patterns. Their work was published in PNAS Nexus and recently referenced in The New York Times. 

“We needed to process and analyze massive datasets of human behavior drawn from millions of posts and notes across different data sources,” said Tan. “This kind of large-scale analysis of data simply wasn’t feasible before. The DSI cluster allowed us to train large language models to detect subtle linguistic patterns in texts. This kind of analysis opens up a new world of possibilities for how AI can help us understand social dynamics, information spread, and human behavior as a whole,” he said, adding, “We couldn’t do this work without it.”

Forging New Paths in Artificial Intelligence and Machine Learning

In addition to enabling interdisciplinary research, the cluster is supporting breakthrough research in the rapidly-growing data science, AI, and machine learning space. 

Victor Veitch, Assistant Professor of Statistics and Data Science, and his team have produced award-winning research on how large language models encode concepts, including work that won the Best Paper Award at ICML 2024. Their papers “Uncovering meanings of embeddings via partial orthogonality” (Jiang et al., NeurIPS 2023) and “On the origins of linear representations in large language models” (Jiang et al., ICML 2024), as well as a more recent presentation at ICLR on “The geometry of categorical and hierarchical concepts in large language models” (Park et al., ICLR 2025 Oral) have already been cited across domains including vision-language models, mechanistic interpretability, and concept engineering.

“The computational demands of modern AI interpretability research are enormous,” said Veitch. “To understand how language models represent concepts, we need to run thousands of experiments across different models and datasets. The cluster lets us conduct rigorous scientific investigations of AI systems at scale, which is critical for developing more reliable and trustworthy AI.”

Ari Holtzman, Assistant Professor of Computer Science and Data Science, is building on his foundational work in neural text generation (his 2019 paper “The Curious Case of Neural Text Generation” is now used in OpenAI’s API) to tackle the next frontier: understanding how AI systems can be aligned with human values and ensuring they behave reliably and safely. 

“Our research investigates what large language models can and can’t do, and what entirely new capabilities emerge when humans work with them,” Holtzman explained. “To understand these complex systems, we need to systematically probe them—through creative prompting, targeted finetuning, model surgery like deleting or swapping components, and deep dives into their internal mechanisms. All this requires massive computational resources, which would be difficult without the cluster’s parallel processing power.”

Tian Li, Assistant Professor of Computer Science and Data Science, uses the cluster to advance techniques that facilitate machine learning across distributed systems while preserving privacy.

“Modern machine learning research increasingly requires experiments across hundreds or thousands of scenarios to ensure robustness,” Li explained. “Being able to do this efficiently has allowed us to quickly iterate on new ideas and validate theoretical insights.”

Especially as artificial intelligence becomes more ubiquitous, UChicago researchers are playing an important role in defining the field.

Preparing the Next Generation

Often, high-performance computing has been accessible only to researchers with substantial grant funding, or access to specialized supercomputing facilities. But the cluster provides the resources for discovery across career stages, and is frequently used by graduate students and postdoctoral researchers.

The cluster also serves as a crucial training resource through programs like the Data Science Clinic. Nick Ross, the Data Science Clinic Director, spoke about the importance of cluster access in training the next generation of data scientists: “When students can iterate quickly on complex models or analyze datasets that would have been impossible to work with just a few years ago it changes how they think about problems and problem solving. They start approaching problems differently, no longer constrained by the limitations of the laptop in front of them. Through programs like the Data Science Clinic, students are graduating with hands-on experience using tools that are becoming increasingly essential across every field.”

The Team Behind The Cluster

To the end user, it may feel like magic–and that’s the idea. But behind the proverbial curtain, installing and maintaining all those servers, is the two-person team of Bennett Hunter and Maria Hernandez.

“The easiest way to understand it,” explained Hunter, “is that the machines we use are able to do math incredibly fast. Decimals are more difficult for computers to process, but our system can perform about two quadrillion 64‑bit floating‑point operations per second, and each 64-bit number carries about 16 decimal digits of precision.”

As the cluster expands, each new compute node needs to be fit into the parallel processing system. “We use our own independent system to configure servers for the cluster. After we test it, we install it in the data center in a subbasement on campus. Then we rack it–” which seemed like the easiest part of the process, until Hunter clarified each piece can weigh up to 150 pounds, adding, “My first assignment was to set up five nodes at once.” Luckily, Hunter was joined earlier this spring by a first team member, Hernandez.

“We’re engineers at heart, so we love to see this kind of state of the art technology,” Hunter said, “but the greatest piece of engineering is one someone can use without an instruction manual, so we’re committed to making the user experience as smooth as possible.”

To this end, they provide users training on how to use the cluster effectively and recently launched a website with guidance on how to get started as well as common uses.

As research becomes increasingly data-driven and computationally intensive, access to high-performance computing will only grow from luxury to necessity. “Often when new faculty join, one of the first questions we’re asked is, ‘How much compute is there?’ because it’s so vital to answering the research questions they’re exploring,” said David Uminsky, Executive Director of the Data Science Institute. “We are building a world class AI research institute, so it’s important to us to empower our researchers with the resources they need to pursue ambitious, high-impact work.”

From fundamental questions about the nature of artificial intelligence to practical applications in neuroscience and climate science, our researchers’ accomplishments demonstrate how strategic investment in a shared computational infrastructure can serve as a force-multiplier, amplifying capabilities and catalyzing discovery across domains at UChicago.