Simulating plasma turbulence presents significant computational challenges due to the complex interplay of multi-scale dynamics. In this work, we investigate the use of convolutional neural networks to improve the efficiency of plasma turbulence simulations, focusing on the Hasegawa–Wakatani model. The networks are trained to learn the closure terms in large eddy simulations, providing a computationally cheaper alternative to the high-resolution numerical solvers for capturing the effects of high-frequency components. This study is the first to successfully apply machine learning to predict plasma behavior for adiabatic coefficients beyond the training range for the Hasegawa–Wakatani equations. We generate ground truth simulations for three values of the adiabatic coefficient (), and train our models on one, or two. The evaluation is then performed on the remaining values. The models generalize well and accurately predict the particle flux up to a factor 5 outside the training range. Finally, we address a key challenge in machine-learning-accelerated plasma simulations—initialization—by starting simulations for previously unseen adiabatic coefficients C with states from existing simulations at other known C values. This approach removes the need for expensive direct numerical simulations for initialization while maintaining physical accuracy and a fast convergence rate. Overall, the results highlight the model's strong generalization capabilities and its potential for accelerating plasma turbulence simulations with more complex sets of parameters.
This paper presents a novel approach to accelerate plasma turbulence simulations using machine learning. The researchers trained convolutional neural networks to learn the closure terms in large eddy simulations, allowing for computationally cheaper predictions of plasma behavior beyond the training range. The models demonstrated strong generalization capabilities, accurately predicting particle flux up to 5 times outside the training range.