Unlocking the Physics of Turbulence with AI

The U.S. Department of Energy has selected Iván Bermejo-Moreno, associate professor at USC Viterbi's Department of Aerospace & Mechanical Engineering, to lead a multi-institutional project funded through its Genesis Mission to apply AI and machine learning to improve turbulence prediction in fluid dynamics. The project, a collaboration with the University of Michigan and Argonne National Laboratory, aims to develop models that learn from the history of flow interactions rather than snapshots, potentially enhancing large eddy simulation accuracy across different flows and simulation codes.

You’re 30,000 feet in the air, and the seatbelt sign comes on. The pilot warns of turbulence ahead, and you regret having just poured that glass of cream soda. What follows – a few minutes of bumps, and jolts and sticky spillages – is an encounter with turbulence at a scale we can physically feel. But imperceptibly small fluid motions are just as consequential, acting to dissipate energy within the flow at large. Engineers simulate turbulent flows to predict how technologies will perform, reducing aerodynamic drag, improving combustion, and controlling the movement of fluids through industrial systems. But calculating every motion, down to the smallest scales, typically outstrips even the capabilities of today’s supercomputers. Iván Bermejo-Moreno https://viterbi.usc.edu/directory/faculty/Bermejo-Moreno/Ivan , associate professor at the USC Viterbi https://viterbischool.usc.edu/ ’s Department of Aerospace & Mechanical Engineering https://ame.usc.edu/ , develops methods to improve the prediction of these smaller-scale dynamics. His research group https://computational-aerospace-laboratory.gitlab.io/ identifies structures within turbulent flows, characterizes their geometry, and analyzes how they interact. The US Department of Energy DOE has selected Bermejo-Moreno to lead a multi-institutional project funded through its Genesis Mission https://www.energy.gov/undersecretaryforscience/genesis-mission/genesis-mission , a national initiative that pairs DOE’s 17 national laboratories with universities and industry to pursue AI-driven scientific discovery. The project is a collaboration between Bermejo-Moreno’s group at USC; Ricardo Vinuesa https://aero.engin.umich.edu/people/ricardo-vinuesa/ ‘s group in the Aerospace Department at the University of Michigan; and computational scientists Ramesh Balakrishnan https://www.anl.gov/profile/ramesh-balakrishnan and Riccardo Balin https://www.alcf.anl.gov/about/people/riccardo-balin , in the Computational Science and Leadership Computing Facility divisions at Argonne National Laboratory. “The bridge here is between artificial intelligence and machine learning methods applied to this analysis of turbulent flow physics,” Bermejo-Moreno said. “We have methodologies that we’ve been developing over several years, and now we want to inject AI and ML techniques to accelerate and enhance those approaches.” Beyond snapshot simulations One of the leading tools for high-fidelity turbulence prediction is large eddy simulation LES , which calculates the larger motions in a flow while using models to approximate the effect of smaller motions. That model’s accuracy is often the limiting factor for the whole simulation, as most such models work from a snapshot of the flow’s current state to estimate how smaller scales behave. Bermejo-Moreno’s research team proposes an alternative approach. Applying the methods practiced in his lab, they plan to identify coherent structures within the flow, map the geometry, and track how each formation interacts as the flow evolves. The result is a model that learns from a history of interactions, not just a snapshot. The team estimates that this added context will generate a model that holds up across different flows and different simulation codes. “This approach can make the analysis pipelines much more effective, as well as generalizing them to other flows that go beyond what we have simulated previously,” said Bermejo-Moreno. The project will run on some of the world’s most powerful supercomputers, combining Argonne’s high-performance computing and AI expertise, Michigan’s work on AI for fluid flows, and USC’s experience in turbulence-structure analysis. “I see this group of researchers as a synergistic team,” Bermejo-Moreno said. “The elements that all of us bring together make us a strong contender to challenge the current state of the art.” Optimizing design Aircraft, turbomachinery and wind turbines are usually designed using physical experiments and lower grade computational methods; the most detailed simulations are too expensive to run repeatedly while comparing designs. The same limitation applies to emerging technologies; Bermejo-Moreno cites “inertial confinement fusion” in particular, one of the methods proposed by the DOE as an alternative means of generating energy. Faster, more reliable models could surpass that limitation, enabling engineers to use high-fidelity simulation while comparing designs, and running enough simulations to test the design as conditions change — the process known as uncertainty quantification. Physical experiments would remain essential, Bermejo-Moreno said. “Experiments tend to be the strongest means of assessing turbulent flows. But they are also expensive, and the diagnostics of experiments are often time-consuming.” This means that numerical simulations need to be as accurate as they are efficient. “By trying to develop these fine-scale models that are more robust, more capable and faster, we can tackle problems that we couldn’t before,” said Bermejo-Moreno. “That’s going to be highly impactful for design optimization and uncertainty quantification in turbomachinery, aerospace and energy systems.” Published on August 24th, 2026 Last updated on August 24th, 2026