🔬 physics

A Neural Operator Closure for Landau Damping in Electrostatic Plasma

This paper introduces a data-driven, non-Markovian Fourier Neural Operator closure trained online within a differentiable fluid solver that successfully reproduces both linear and nonlinear Landau damping in electrostatic plasma by learning a memory-dependent moment-to-flux relation from simulation trajectories rather than kinetic snapshots.

Original authors: Samuel Burles, Enrico Camporeale, Oreste Pezzi

Published 2026-08-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Samuel Burles, Enrico Camporeale, Oreste Pezzi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Invisible Tug-of-War in the Sky

Imagine the air around us isn't just empty space, but a chaotic, invisible ocean made of tiny, super-fast charged particles called plasma. This stuff is everywhere: in the sun, in lightning, and even in the neon lights of a city. Scientists have two main ways to study this wild ocean. The first is like trying to track every single fish in the sea individually; it's incredibly accurate but requires so much computing power that it's often impossible to do for big systems. The second way is to treat the ocean like a smooth, flowing river, ignoring the individual fish and just looking at the water's average speed and pressure. This is much faster, but it's like trying to predict a storm by only looking at the water's surface while ignoring the wind and waves underneath.

The big problem is that sometimes, the "invisible fish" (the individual particles) do things that the "smooth water" (the average flow) can't predict. One famous example is called Landau damping. It's a bit like a surfer on a wave: if the wave is just right, the surfer can catch it and speed up, stealing energy from the wave until the wave itself disappears. In plasma, particles can do this to electric waves, making them vanish without any friction. For decades, scientists have struggled to write a simple set of rules (a "closure") that lets the fast, smooth river model predict when and how these waves will vanish, especially when the waves get big and chaotic. Without a good rule, the smooth model either gets the physics wrong or crashes into a digital wall.

The Smart "Memory" Closure

In this paper, the authors, Samuel Burles and his team, tried to teach a computer to write those missing rules using a special kind of artificial intelligence called a Neural Operator. Instead of just memorizing a list of answers, they trained this AI to act as a "closure"—a bridge that tells the fast fluid model what to do next based on what's happening right now and what happened a little while ago.

Here is the clever twist: most AI models for physics are trained by looking at a snapshot of the "perfect" answer (from the slow, expensive simulation) and trying to copy it. The authors did something different. They built their AI directly inside the fast fluid simulator and let it learn by watching the entire story of the simulation play out. They asked the AI: "If you make this prediction, does the whole simulation stay stable and look like the real physics?" This is like teaching a driver not just by showing them a photo of a perfect turn, but by letting them drive a car and only giving them a high score if they don't crash over a long, winding road.

What they found:

It works for both calm and chaotic waves: They trained the AI on small, gentle waves (linear) and huge, wild waves where particles get trapped and bounce around (nonlinear). A single AI model learned to handle both regimes perfectly, predicting how the waves would fade away or bounce, matching the expensive "perfect" simulations almost exactly.
It's a "smart guesser," not a copycat: In the wild, chaotic part of the simulation, the AI's prediction for the "heat flux" (the flow of energy) didn't match the perfect simulation point-for-point. Instead, it learned a different kind of heat flux that acted as a "magic fix." It compensated for the fact that the fast fluid model was ignoring the tiny, hidden details of the particles. It was like a chef who doesn't have the exact spice but knows exactly how much salt to add to make the dish taste right.
It has a memory: The AI wasn't just looking at the current moment; it looked at a "trailing window" of the last 50 steps of history. This is crucial because the physics of Landau damping relies on the history of how particles have been moving. The AI learned to use this memory to predict the future, and when the researchers tested it on waves it had never seen before (different sizes of waves), it generalized beautifully, interpolating between the sizes it knew.

What it isn't:
The authors are careful to say this isn't a magic bullet that solves every plasma problem yet. They tested it specifically on one-dimensional, electric-only waves. They also noted that the "magic fix" the AI learned for the heat flux depends on the specific computer code they used; if you change the code, the AI might need to relearn its trick. However, they proved that training the AI inside the simulator (online) makes it much more stable and reliable than the old method of training it on static data (offline), which often led to the simulation crashing.

In short, the paper shows that by letting an AI learn from the consequences of its own predictions in a fast simulation, we can build a model that captures the complex, invisible dance of plasma particles without needing to track every single one. It's a promising step toward simulating space weather and fusion energy with the speed of a fluid model but the accuracy of a particle model.

Problem Statement
Collisionless plasma dynamics are inherently multi-scale, where kinetic effects at the particle level significantly influence macroscopic fluid behavior. While kinetic simulations (solving the Vlasov equation) capture these effects accurately, they are computationally prohibitive for large-scale systems due to the high dimensionality of phase space. Conversely, fluid models, derived by taking velocity moments of the Vlasov equation, are computationally efficient but require a "closure" relation to truncate the infinite hierarchy of moment equations. Traditional analytic closures, such as the Hammett-Perkins model, are derived from linear kinetic theory and often fail to capture nonlinear phenomena like particle trapping or to generalize across varying initial conditions. Recent machine learning approaches have attempted to learn these closures, but many rely on "offline" (a priori) training, where models are fitted to reproduce kinetic heat fluxes on pre-computed data. This approach often yields models that are unstable when deployed in independent fluid solvers, as small pointwise errors can compound over time. Furthermore, existing studies often demonstrate accuracy only for specific initial conditions rather than generalizing across a range of amplitudes spanning linear to nonlinear regimes.

Methodology
The authors propose a data-driven, non-Markovian closure for one-dimensional electrostatic Landau damping using a Fourier Neural Operator (FNO). The core innovation lies in the training strategy:

Online Training (A Posteriori): Unlike offline methods, the FNO is embedded directly within a differentiable fluid solver (implemented in PyTorch). The loss function is computed on the trajectories of the resolved fluid moments (density, velocity, pressure) produced by the closed simulation, rather than on individual kinetic snapshots or the kinetic heat flux itself. This ensures the closure is optimized for numerical stability and consistency with the specific fluid solver dynamics.
Non-Markovian Architecture: The closure acts on a trailing window of the recent history of resolved moments (specifically $K=50$ timesteps) rather than just the instantaneous state. This design is motivated by the Mori-Zwanzig formalism, which suggests that projecting out unresolved dynamics introduces a memory integral. The memory window allows the network to capture the phase-mixing history essential for Landau damping.
Input Conditioning and Scaling: To handle the vast dynamic range between linear and nonlinear regimes, the network inputs are normalized by a characteristic amplitude ( $A_{scale}$ ) derived from the density perturbation. An additional channel carrying $\log A_{scale}$ conditions the network, allowing a single model to modulate its behavior across different perturbation amplitudes.
Target Variable: The network outputs the heat-flux gradient ( $\partial_x q$ ) directly, rather than the heat flux itself, to avoid derivative approximation errors and to preserve the ability to reproduce net heating effects associated with Landau damping.

Key Contributions

Stable Online Training: The paper demonstrates that training a neural closure online within a differentiable solver yields a model that is numerically stable when deployed in independent simulations, addressing the instability issues common in offline-trained closures.
Generalization Across Regimes: A single FNO model, trained on a distribution of initial perturbation amplitudes, successfully reproduces both linear and nonlinear Landau damping. It generalizes to initial amplitudes outside the training set (interpolation) and remains stable, whereas extrapolation to amplitudes significantly larger than the training maximum shows limitations.
Effective vs. Kinetic Flux: The study establishes that the learned closure does not necessarily reproduce the kinetic heat flux pointwise. Instead, it learns an "effective" heat flux that compensates for the truncation of higher-order moments, ensuring the correct evolution of the resolved fluid moments.
Physical Interpretation: Through sensitivity analysis (ablation tests and Hilbert-Schmidt Independence Criterion), the authors show the closure computes a genuine relation between plasma state and heat flux, relying on a structured memory window that captures the oscillation phase and becomes increasingly dependent on heat-flux history as the system enters the nonlinear trapping regime.

Results

Linear Regime: The online-trained FNO reproduces the linear damping rate and oscillation phase with high fidelity (error $< 0.2\%$ in damping rate), outperforming an offline-trained FNO (which failed to capture oscillations) and matching the performance of a calibrated Hammett-Perkins closure.
Nonlinear Regime: For large initial amplitudes ( $A=10^{-1}$ ), the online FNO captures the initial damping, saturation, and bounce oscillations of the trapped state over long simulation times ( $500 \omega_p^{-1}$ ). In contrast, the offline-trained FNO and the Hammett-Perkins closure fail to reproduce the nonlinear saturation and modulation.
Generalization: A model trained on amplitudes ranging from $10^{-3}$ to $10^{-1}$ accurately predicts damping rates and saturation energies for test amplitudes within this range. The error in resolved moments (density, pressure, velocity) remains low ( $10^{-2}$ to $10^{-1}$ ) even in the nonlinear regime, although the error in the heat-flux gradient itself grows to order unity, confirming the "effective flux" hypothesis.
Energy Conservation: While not explicitly constrained, the learned closure exhibits bounded energy drift comparable to the kinetic reference in linear regimes and bounded (though larger) drift in nonlinear regimes, indicating it does not introduce significant spurious dissipation or injection.

Significance
The paper claims to provide a concrete step toward broadly applicable machine-learned plasma closures. By demonstrating that a single online-trained model can generalize across linear and nonlinear regimes while remaining numerically stable in independent fluid simulations, the work addresses critical gaps in the current literature regarding stability and generalization. The authors position this as a viable route for performing fluid-cost simulations that retain essential kinetic effects, provided the training data spans the dynamically relevant regimes. The study emphasizes that the learned closure acts as an effective operator compensating for truncated moments rather than a direct reconstruction of kinetic variables, a distinction that is crucial for the stability and physical consistency of large-scale fluid modeling.

The Invisible Tug-of-War in the Sky

The Smart "Memory" Closure

More like this