Scientists at Los Alamos National Laboratory have cracked a problem that has stumped physicists for generations: predicting how individual particles behave when caught in the chaos of turbulent flows. From dust swirling in a tornado to sugar grains spinning in a cup of coffee, particles in turbulence follow paths so complex they've resisted mathematical modeling at meaningful scales—until now.
The breakthrough, published in the Proceedings of the National Academy of Sciences, represents the first data-driven machine learning framework capable of capturing particle dynamics in turbulent systems with both short-term precision and long-term statistical accuracy. Led by Los Alamos scientist Daniel Livescu and lead author Xander de Wit, the research team applied neural networks layered onto a mathematical framework called the Mori-Zwanzig formalism, which breaks complex dynamical systems into resolved components based on current observations and past history—what physicists call "memory effects."
"Modeling turbulence is a big, open problem, and it's probably the hardest problem in classical physics," Livescu said. "A subset of that challenge is modeling particle motions within turbulence. To meet that challenge, our artificial intelligence approach offers an innovative theoretical construct tested with a real-world application."
What makes this approach revolutionary is its ability to work at scale. Turbulence is a multiscale phenomenon—characterized by enormous swirling vortexes that cascade down into progressively smaller and smaller eddies. At these smaller scales, predicting particle trajectories and velocities becomes nearly impossible using conventional methods. The new machine learning model trained on short-term predictions achieves something that has long eluded researchers: accurate statistical behavior of Lagrangian turbulence over longer time spans without the staggering computational cost of traditional simulations.
The team built what physicists call a "surrogate model"—essentially a learned replica of turbulent dynamics that captures all the system's behavior while remaining computationally efficient. This matters far beyond abstract curiosity. Turbulent flows are central to understanding weather systems, astrophysical phenomena, and even the controlled fusion reactions needed for advanced energy production. Better models of particle dynamics within turbulence could reshape how scientists approach these complex problems.
The research represents a convergence of artificial intelligence and physics-aware mathematics. The Mori-Zwanzig formalism proved critical because it accounts for memory—the way a particle's current behavior depends not just on its immediate surroundings but on its entire history. This memory-inclusive approach is what allows the model to make accurate predictions across different time scales.
De Wit, who authored the paper while conducting research in Europe, will join Los Alamos full-time next fall as a Richard Feynman Distinguished Postdoctoral Fellow. He suggests the implications extend well beyond turbulence in classical fluids. "Future problems might be applied to things like crowd movements, where many of the same Lagrangian aspects are also at work," de Wit said. "Providing models of those types of situations will be a logical and useful extension of this work."
As machine learning continues to sharpen our ability to model chaotic systems, this framework opens a new frontier in studying phenomena that seemed locked away by the inherent unpredictability of chaos itself.