Imagine a computer so powerful it can perform more calculations in a second than a billion people could complete working nonstop for decades.
That's the scale of Discovery, a next-generation supercomputer expected to become the fastest in the United States - and potentially the world - when it officially launches in 2029 at the U.S. Department of Energy's Oak Ridge National Laboratory.
Before it can drive breakthroughs in energy, security, manufacturing and human health, however, Discovery must run a gauntlet of tests.
A University of Delaware team led by Sunita Chandrasekaran, the David L. and Beverly J.C. Mills Career Development Chair in Computer and Information Sciences and director of the First State AI Institute, is one of nine groups selected nationwide through a highly competitive process to stress-test the system and prepare it for scientific use.
"Our goal is to run thousands of experiments across many different physics scenarios and see what breaks - whether it's software bugs, performance bottlenecks or system issues under stress," Chandrasekaran said. "That's where we learn, and where our team can help fix problems in real time, working directly with the system's builders."
The work is part of ORNL's Discovery Center for Accelerated Application Readiness (CAAR), which is preparing cutting-edge scientific applications under the U.S. Department of Energy's Genesis Mission. The initiative aims to build the world's most powerful scientific platform by combining AI, high-performance computing and emerging quantum technologies to tackle complex problems at unprecedented speed.
Pressing Peak Performance
Discovery goes beyond today's exascale machines, which exceed a quintillion calculations per second. But raw speed isn't enough. The system must sustain that performance across complex, AI-driven workloads.
Testing will push every component - processors, memory, networking and data movement - to uncover weaknesses and ensure reliability.
"You're not just asking, 'Does it run-'" Chandrasekaran said. "You're asking, 'Can it handle the most demanding science continuously, at scale, without failing-'"
Once the initial computing cluster is installed later this year, her team will run thousands of simulations using Particle-in-Cell on GPU (PIConGPU), a software package that models how charged particles, such as electrons and ions, interact in electromagnetic fields.
By using specialized chips called graphics processing units (GPUs) to simulate billions of particles simultaneously, the system will allow researchers to study plasma behavior, with applications in fusion energy, spacecraft propulsion, astrophysics and laser-based medical technologies.
A key challenge in fusion energy - despite its promise as a clean, nearly limitless energy source - is efficiently transferring laser energy into the tiny fuel pellets that drive reactions.
To address this, the team will leverage Discovery's enhanced performance and memory to pair high-fidelity PIConGPU simulations with AI, identifying optimal fusion target materials and designs.
A Rare Opportunity for Students
The project gives students hands-on experience with next-generation computing.
Nikhil Rao, a doctoral student in computer science at UD, is developing workflows that connect plasma simulations with machine learning models, without storing the massive data they generate.
"We're working with simulations that produce data at incredible rates, on the order of petabytes per second," Rao said. A petabyte - about 1 million gigabytes - is enough to store 250 million photos. "Instead of storing it, we send the data directly to machine learning models while it's still in memory."
Rao said the opportunity to work on a system still in development is rare.
"Very few people get to work directly with the hardware vendors and scientists building a system like this," he said. "It's quite amazing."
Rao will collaborate with engineers at AMD and Hewlett Packard Enterprise, along with researchers at Oak Ridge National Laboratory and Helmholtz-Zentrum Dresden-Rossendorf (HZDR) in Germany, a key collaborator on Chandrasekaran's large-scale computing efforts.
HZDR previously partnered with her to stress-test Frontier, Oak Ridge's current exascale system, and now brings expertise in plasma physics and high-performance computing to help design demanding, scientifically meaningful test scenarios for Discovery, which is expected to surpass Frontier in both speed and efficiency.
"This work builds on a longstanding collaboration with Sunita and her team," said Michael Bussmann, plasma physicist and founding manager of the Center for Advanced Systems Understanding at HZDR. "Their expertise in high-performance computing and scalable AI has been critical to our participation in this prestigious program. Our success relies on continued international collaboration like this."
For the United States, Discovery represents a major step forward in high-performance computing and AI. For UD researchers and students, it offers a front-row seat to the future.
"We're preparing for a machine that doesn't fully exist yet," Chandrasekaran said. "That's what makes this work so fascinating and rewarding. We're not just preparing for what's next - we're helping make it possible."