[Paper Review] Physics-Preserving AI-Accelerated Simulations of Plasma Turbulence

Robin Greif, F. Jenko+1 more|arXiv (Cornell University)|Sep 28, 2023

Model Reduction and Neural NetworksPhysics and Astronomy3 citations

TL;DR

This paper proposes a physics-preserving machine learning approach that combines Large Eddy Simulation (LES) with a neural network-based sub-grid-scale (SGS) model to accelerate plasma turbulence simulations. By resolving only large-scale dynamics and using a learned correction on resolved fields, the method reduces computational cost by three orders of magnitude while preserving key statistical properties of the turbulent system, including fluxes, energy, and spectral distributions.

ABSTRACT

Turbulence in fluids, gases, and plasmas remains an open problem of both practical and fundamental importance. Its irreducible complexity usually cannot be tackled computationally in a brute-force style. Here, we combine Large Eddy Simulation (LES) techniques with Machine Learning (ML) to retain only the largest dynamics explicitly, while small-scale dynamics are described by an ML-based sub-grid-scale model. Applying this novel approach to self-driven plasma turbulence allows us to remove large parts of the inertial range, reducing the computational effort by about three orders of magnitude, while retaining the statistical physical properties of the turbulent system.

Motivation & Objective

• Address the computational intractability of simulating turbulent plasma dynamics at high resolution.
• Overcome the limitations of traditional LES in handling large parts of the inertial range, where energy cascades occur.
• Develop a machine learning-based sub-grid-scale model that preserves physical consistency while drastically reducing computational cost.
• Enable accurate, fast simulations of self-driven plasma turbulence relevant to magnetic confinement fusion energy research.
• Demonstrate that a hybrid numerical-ML approach can maintain statistical equilibrium properties even when removing most of the inertial range.

Proposed method

• Apply Large Eddy Simulation (LES) to resolve only the largest-scale dynamics of plasma turbulence, filtering out small-scale motions.
• Introduce a neural network-based sub-grid-scale (SGS) model that learns corrections to the resolved fields using data from high-resolution direct numerical simulations (DNS).
• Use a non-propagated field (the electrostatic potential φ) as input to the neural network, enabling effective modeling of unresolved dynamics.
• Train the neural network to predict the effects of unresolved scales on the resolved fields (n and Ω) via a loss function minimizing differences in flux, energy, and other statistical metrics.
• Implement a hybrid simulation framework where the resolved equations are solved numerically, and the SGS model provides closure for sub-grid dynamics.
• Apply downsampling to the DNS data to create training data for the ML model, ensuring the model learns from realistic turbulent states at reduced resolution.

Experimental results

Research questions

• RQ1 Can a machine learning-based SGS model preserve the statistical properties of plasma turbulence when large parts of the inertial range are removed?
• RQ2 To what extent can computational cost be reduced in plasma turbulence simulations without sacrificing physical fidelity?
• RQ3 How well does a neural network-based SGS model reproduce key observables such as turbulent particle flux (Γₙ), energy (E), and potential spectra compared to direct numerical simulation?
• RQ4 Can the ML model generalize across different resolutions and parameters (e.g., ν, c₁) while maintaining consistency with the underlying physics?
• RQ5 Does the inclusion of learned corrections on resolved fields outperform previous ML models that directly predict field variables or use simpler closures?

Key findings

• The proposed ML-accelerated simulation reduces computational cost by approximately three orders of magnitude (1,000x speedup) compared to direct numerical simulation at 512×512 resolution.
• The model preserves the mean turbulent particle flux (Γₙ) within 0.57 ± 0.01 (vs. 0.55 ± 0.00 for DNS), showing near-identical statistical behavior.
• Energy (E) and current flux (Γ_c) are reproduced with high accuracy: E mean = 0.450 ± 0.056 (model) vs. 0.450 ± 0.056 (DNS) at 64×64 resolution.
• Spectral distributions of phase shifts (δ(k_y)) align closely with DNS results within one standard deviation across the retained scale range.
• Distribution shapes of all physical metrics (e.g., ∂ₜE, Γₙ) are preserved between the model and downsampled reference data, with only minor center shifts that vanish under longer temporal unrolling.
• The model outperforms previous approaches: for example, at 32×32 resolution, the model achieves a flux of 0.57 ± 0.01, while a prior method yields 0.84 ± 0.01, indicating superior accuracy and stability.