90,000x Faster: Breakthrough Cuts Rocket Engine Simulations from Days to Seconds

Developing methods that turn three days of super‑computer crunching into mere seconds, a University of Texas at Austin led team has created reduced-order models that slash rocket‑engine simulation time by an astonishing  90,000x, opening the door to faster, smarter design of next‑generation propulsion systems.

The work targets cutting‑edge rotating detonation rocket engines (RDREs), which ignite fuel with spinning detonation waves. RDREs could outperform the propulsion devices used in today’s rockets, but their complex combustion physics makes computation painfully slow. “For example, a single simulation of one RDRE design over one millisecond of operation can take three days on a supercomputer,” explains lead author Ionut‑Gabriel Farcas. Such wait‑times cripple engineers’ ability to simulate, test and refine new ideas.

Farcas is a former postdoctoral fellow at the Oden Institute for Computational Engineering and Sciences at UT who conducted research with the Willcox Research Group and the Center for Scientific Machine Learning and is now an assistant professor in the Department of Mathematics at Virginia Tech. Farcas developed the solution in collaboration with Professor Karen Willcox and their colleagues from the Air Force Research Laboratory (AFRL). Their work, titled “Distributed computing for physics-based data-driven reduced modeling at scale: Application to a rotating detonation rocket engine,” was recently published in Computer Physics Communications, as part of the special issue on Advances in Physics-aware Machine Learning. 

They used scientific machine‑learning techniques, training surrogate models on data from a high-fidelity rocket engine simulation. The dataset comprised flow fields such as pressure, velocity, specific volume, chemical mass fractions, and temperature. Willcox, a leading expert in model reduction and uncertainty quantification, brought her expertise in reduced-order modeling to the project. Farcas’ key innovation - distributed Operator Inference - farms this training across thousands of cores on parallel computers, producing predictive surrogates in a matter of seconds even for large-scale datasets.

This research builds on growing efforts by the U.S. Air Force to accelerate RDRE design through advanced simulation tools. AFRL has been developing computational models to support propulsion system design and guide experimental campaigns. While these simulations offer critical insights, they’re notoriously expensive, requiring enormous compute resources for even microsecond-level analyses. Therefore, bridging this gap between insight and efficiency has become a major priority for AFRL.

The result: engineers can iterate through “what‑if” scenarios almost instantly instead of waiting days, paving the way for real‑time design exploration and optimization.

"Our collaboration with the Air Force Research Laboratory has been a fantastic opportunity to push the boundaries of reduced-order modeling for complex systems. Rotating detonation rocket engines bring a combination of physics complexity and data sparsity, making it a major challenge to achieve truly predictive reduced-order models." said Willcox, who is also the Director of the Oden Institute.  

The project was powered by a unique combination of domain expertise and high-performance computing. Farcas credits AFRL for providing access to realistic RDRE simulation data and domain expertise, the Frontera supercomputer at the Texas Advanced Computing Center (TACC) at UT Austin for enabling efficient surrogate training in parallel across 2,048 cores, and Willcox for her guidance and mentorship throughout the research process. “AFRL’s expertise, Karen’s guidance, and TACC’s computing power all came together to make it happen,” he said. “It’s a reminder that research is a team sport.”

Farcas views the project as a prime example of scientific machine learning applied to real‑world engineering. “From the beginning, I’ve loved doing computational mathematics for real‑world applications, and there’s nothing more real‑world than helping design better rockets!” he noted. Just as important, he urges emerging researchers to keep an open mind: “Stay curious. In interdisciplinary research, you often find solutions in unexpected places.”

Acknowledgements

This work was supported in part by AFRL Grant FA9300-22-1-0001 and the Air Force Center of Excellence on Multifidelity Modeling of Rocket Combustor Dynamics under grant FA9550-17-1-0195.