When aerospace vehicles travel at hypersonic speeds, or materials face temperatures too high to replicate in a physical experimental lab, how can engineers be sure their designs will hold up?
The University of Utah’s John and Marcia Price College of Engineering is now joining a $16 million effort to address this challenge.
The U.S. National Nuclear Security Administration, an agency within the Department of Energy, has announced a grant to establish the SAGEST Predictive Simulation Center, a multi-university collaboration headquartered at the University of Virginia.
SAGEST — short for Stochastic Simulations of Ablative Geometries with Error-Learning in Space and Time — will develop artificial-intelligence-powered simulation tools that give scientists confidence in exploring extreme physical conditions that are too difficult or costly to test directly.
Mary Hall, director and professor at the Kahlert School of Computing, will represent Utah in the collaboration. Her role will be to develop code-generation technologies for working with the Tuolomne supercomputer at Lawrence Livermore National Laboratory. Tuolomne is a sibling system to El Capitan — the #1 supercomputer in the Top500 — and uses the same AMD MI300 GPU architecture.
New code-generation technologies are necessary because the central innovation of the SAGEST project lies in how it layers different levels of computational precision, balancing accuracy and efficiency in its predictions while quantifying uncertainty.
In the context of hypersonic vehicles, there is a need to simulate physical interactions at multiple different scales. Where precision is paramount, high-fidelity solvers might simulate the underlying physics of individual molecules. Such precision is costly, however, and can’t be used to get a big-picture view of the vehicle’s overall aerodynamic properties. For those more general simulations, low-fidelity solvers are used.
The challenge is to have the results of these differently-scaled simulations talk to one another as a cohesive whole. The high-fidelity and low-fidelity solvers must exchange information continuously if the simulation is to adapt in real time and update predictions across the entire system.
With SAGEST’s approach, classical numerical algorithms are augmented by artificial intelligence to minimize errors in the process. That, in turn, requires a new way of interfacing with the supercomputer that has the computational power to run these simulations.
“The way Tuolomne handles memory changes how we move data through the system,” says Hall. “Our job is to optimize this data movement within the supercomputer, as we go from computations in finer grain models up to coarser ones.”
SAGEST will be headquartered at UVA and led by Xinfeng Gao, professor of mechanical and aerospace engineering at UVA’s School of Engineering and Applied Science. While the center’s immediate focus includes aerospace use cases, its simulation platform can be applied to any field where physical experiments fall short. This includes complex energy systems, advanced manufacturing, materials development and biomedical modeling.
“The core goal of this center is to build simulations you can trust,” Gao said. “We want to develop predictive technologies that don’t just generate results but give scientists and engineers confidence in those results, particularly under extreme conditions.”