Reconstructing High-fidelity Plasma Turbulence with Data-driven Tuning of Diffusion in Low Resolution Grids

2025/09/15 by Kunpeng Li, Youngwoo Cho, Li, Kunpeng +14
Engineering · Earth and Planetary Sciences · Physics and Astronomy · #Fluid Dynamics and Turbulent Flows #Meteorological Phenomena and Simulations #Magnetic confinement fusion research

paper · pdf · doi:10.48550/arxiv.2509.11576

Abstract

Developing physically consistent closure models is a longstanding challenge in simulating plasma turbulence, even in minimal systems such as the two-field Hasegawa-Wakatani (HW) model, which captures essential features of drift-wave turbulence with a reduced set of variables. In this work, we leverage theoretical insights from Direct Interaction Approximation (DIA) to construct a six-term closure structure that captures the dominant turbulent transport processes, including both diffusion and hyper-diffusion. While the mathematical form of the closure is fully prescribed by DIA, the corresponding transport coefficients are learned from data using physics-informed neural networks (PINNs). The resulting Extended HW model with Closure (EHW-C) model reveals several nontrivial features of plasma turbulence: notably, some inferred coefficients become negative in certain regimes, indicating inverse transport, a phenomenon absent in conventional closure models. Moreover, the EHW-C model accurately reproduces the spectral and flux characteristics of high-resolution Direct Numerical Simulations (DNS), while requiring only one-eighth the spatial resolution per direction, yielding a tenfold speed-up. This work demonstrates how theory-guided machine learning can both enhance computational efficiency and uncover emergent transport mechanisms in strongly nonlinear plasma systems.

Citations

• Inverse Physics-Informed Neural Networks for transport models in porous materials
• A generative machine learning surrogate model of plasma turbulence
• Inferring Parameters and Reconstruction of Two-Dimensional Turbulent Flows with Physics-Informed Neural Networks
• Physics-informed machine learning
• PredRNN: A Recurrent Neural Network for Spatiotemporal Predictive Learning
• Reduced operator inference for nonlinear partial differential equations
• Thermodynamically consistent physics-informed neural networks for hyperbolic systems
• PyTorch: An Imperative Style, High-Performance Deep Learning Library
• Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations
• A Style-Based Generator Architecture for Generative Adversarial Networks
• Deep Learning for Precipitation Nowcasting: A Benchmark and A New Model
• Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting
• Inertial Ranges in Two-Dimensional Turbulence

Related

• A generative machine learning surrogate model of plasma turbulence
• Extracting a stochastic model for predator-prey dynamic of turbulence and zonal flows with limited data
• Physics-Preserving AI-Accelerated Simulations of Plasma Turbulence
• Neural networks for turbulent transport prediction in a simplified model of tokamak plasmas
• Surrogate modeling of drift-reduced Braginskii turbulence with resistivity-conditioned Koopman neural operators