Somewhere in a wind tunnel right now, a particle is doing something nobody can predict. Not because physics is broken, but because tracking every molecule of air shoving that particle around would require more computing power than exists on Earth. So researchers did what humans always do when faced with the impossible: they cheated, brilliantly.
A team from Los Alamos National Laboratory, Eindhoven University of Technology, and the University of Twente just published a paper that sounds like a philosophy riddle: "Is memory all you need?" Spoiler alert - for simulating particles tumbling through turbulent flows, the answer is a resounding "actually, yeah, kind of."
The Problem With Pretending Air Doesn't Exist
Here's the thing about turbulence: it's chaos wearing a lab coat. When you drop a leaf into a stream or watch smoke curl from a candle, you're witnessing what physicists call a "highly nontrivial statistical behavior" - which is academic speak for "good luck predicting that."
The gold standard for simulating turbulence is called Direct Numerical Simulation (DNS), and it's exactly as exhausting as it sounds. DNS solves every single equation for every single swirl of fluid, from the building-sized eddies down to the tiniest Kolmogorov-scale ripples where viscosity finally wins. For industrial-scale flows, this requires computational grids with trillions of points and supercomputer time measured in decades.
So what if you just... didn't? What if you could track a particle through turbulence without simulating all that turbulence?
The Ghost of Turbulence Past
Enter the Mori-Zwanzig formalism, a mathematical framework from statistical physics that sounds intimidating but has a simple premise: if you can't see everything happening around you, at least remember what happened before.
The idea is that when you strip a complex system down to just a few observable variables, you lose information. The particle doesn't know about the thousands of invisible vortices buffeting it. But here's the clever bit - the history of that particle's journey contains echoes of those hidden influences. The formalism mathematically separates what you can observe now from what you need to remember from the past.
Think of it like navigating a city blindfolded. You can't see the streets, but if you remember every turn you've made and how bumpy the road was, you can start to infer the neighborhood. The Mori-Zwanzig approach gives that memory a rigorous mathematical structure - a "memory kernel" that captures how the past shapes the present.
Training the Model to Have Trust Issues
The researchers built a machine learning model that learns this memory structure directly from DNS data. Train it on short-term accuracy - can you predict where the particle goes in the next fraction of a Kolmogorov time unit? - and something remarkable happens. The model doesn't just nail the short game; it stays stable and statistically accurate over long time periods too.
This is harder than it sounds. Many surrogate models face a nasty tradeoff: optimize for short-term precision and the errors compound until your simulation veers into nonsense. But by respecting the Mori-Zwanzig structure and training on point-wise error metrics, the team found a sweet spot where short-term accuracy translates to long-term statistical fidelity.
The approach combines the formalism with Takens time-delay embedding, another mathematical tool that reconstructs hidden dynamics from observable sequences. It's like giving the model two different flavors of memory and letting them complement each other.
Why Should Anyone Care About Floating Particles?
Beyond the elegant physics, there are surprisingly practical reasons to care about Lagrangian particle dynamics - "Lagrangian" being the fancy term for following individual particles rather than staring at fixed points in space.
Atmospheric scientists use these models to track pollutant dispersion in cities. Ocean researchers model how plastics and nutrients spread through currents. Combustion engineers track fuel droplets in engines. And increasingly, roboticists are interested in guiding active agents through turbulent environments - whether that's microrobots navigating blood vessels or drones staying stable in gusty conditions.
The paper specifically points to "control of active Lagrangian agents in turbulence" as a killer application. Model predictive control requires rapid evaluation of how the system responds to inputs. Running full DNS in real-time is absurd. Running a trained surrogate? Now you're talking.
The Memory Palace of Machine Learning
What's intellectually satisfying about this work is how it bridges two worlds. The Mori-Zwanzig formalism isn't some dusty theoretical curiosity - it's been around since the 1960s, rooted in the same statistical mechanics that explains why your coffee cools down. But applying it to turbulence has been notoriously difficult because nobody knew how to actually compute those memory operators in practice.
Machine learning offers a path forward. Instead of deriving the memory kernel from first principles (good luck), you learn it from data. The neural network becomes an empirical approximation of something physicists have understood conceptually for decades but couldn't implement at scale.
This pattern - using ML to make classical physics tractable rather than replacing it - represents one of the more promising directions in scientific machine learning. The physics provides structure and interpretability; the data provides the details.
The Fine Print
The researchers tested their approach on homogeneous isotropic turbulence, which is physics-speak for turbulence without walls or preferred directions. Real-world turbulence near surfaces, around obstacles, or in stratified fluids adds complications. The team also focuses on single particles - tracking how particles cluster, collide, and interact adds another layer of complexity.
Still, the core insight translates: if you respect the mathematical structure of how memory enters reduced dynamical systems, machine learning becomes dramatically more effective.
Whether you're trying to predict where a microplastic ends up in the ocean or keep a microdrone stable in someone's bloodstream, understanding how particles remember their turbulent past turns out to be surprisingly useful - even if they don't have brains to remember with.
References:
-
de Wit XM, Gabbana A, Woodward M, Lin YT, Toschi F, Livescu D. Data-driven Mori-Zwanzig modeling of Lagrangian particle dynamics in turbulent flows. PNAS. 2025. DOI: 10.1073/pnas.2525390123 | arXiv:2507.16058
-
Li Y, Yang GY, Petit TL. Synthetic Lagrangian turbulence by generative diffusion models. Nature Machine Intelligence. 2024. DOI: 10.1038/s42256-024-00810-0
-
Lin YT, Tian Y. Data-driven learning for the Mori - Zwanzig formalism: a generalization of the Koopman learning framework. arXiv:2101.05873
-
Parish EJ, Duraisamy K. Data-driven learning of Mori - Zwanzig operators for isotropic turbulence. Physics of Fluids. 2021;33(12):125118. DOI: 10.1063/5.0064108
Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.