Abstract:We designed a new artificial neural network by modifying the neural ordinary differential equation (NODE) framework to successfully predict the time evolution of the 2D mode profile in both the linear growth and nonlinear saturated stages. Starting from the magnetohydrodynamic (MHD) equations, simplifying assumptions were applied based on physical properties and symmetry considerations of the energetic-particle-driven geodesic acoustic mode (EGAM) to reduce complexity. Our approach embeds physical laws directly into the neural network architecture by exposing latent differential states, enabling the model to capture complex features in the nonlinear saturated stage that are difficult to describe analytically, and thus, the new artificial neural network is named as ExpNODE (Exposed latent state Neural ODE). ExpNODE was evaluated using a data set generated from first-principles simulations of the EGAM instability, focusing on the pre-saturated stage and the nonlinear saturated stage where the mode properties are most complex. Compared to state-of-the-art models such as ConvLSTM, ExpNODE with physical information not only achieved lower test loss but also converged faster during training. Specifically, it outperformed ConvLSTM method in both the 20-step and 40-step prediction horizons, demonstrating superior accuracy and efficiency. Additionally, the model exhibited strong generalization capabilities, accurately predicting mode profiles outside the training data set. Visual comparisons between model predictions and ground truth data showed that ExpNODE with physical information closely captured detailed features and asymmetries inherent in the EGAM dynamics that were not adequately captured by other models. These results suggest that integrating physical knowledge into neural ODE frameworks enhances their performance, and provides a powerful tool for modeling complex plasma phenomena.

Neural network classification of eigenmodes in the magnetohydrodynamic spectroscopy code Legolas

Neural network classification of eigenmodes in the magnetohydrodynamic spectroscopy code Legolas

Legolas: magnetohydrodynamic spectroscopy with viscosity and Hall current

A fast neural network surrogate model for the eigenvalues of QuaLiKiz

Prediction of Mode Structure Using A Novel Physics-Embedded Neural ODE Method

A novel unsupervised machine learning algorithm for automatic Alfvénic activity detection in the TJ-II stellarator

Explainable deep learning for the analysis of MHD spectrograms in nuclear fusion

Legolas 2.0: Improvements and extensions to an MHD spectroscopic framework

MHD spectrogram contribution to disruption prediction using Convolutional Neural Networks

Do Androids Dream of Magnetic Fields? Using Neural Networks to Interpret the Turbulent Interstellar Medium

Magnetohydrodynamic spectroscopy of a non-adiabatic solar atmosphere

Classifying different types of solar wind plasma with uncertainty estimations using machine learning

Artificial neural network subgrid models of 2-D compressible magnetohydrodynamic turbulence

Classifying Different Types of Solar-Wind Plasma with Uncertainty Estimations Using Machine Learning

Neural ODEs as a discovery tool to characterize the structure of the hot galactic wind of M82

Using Convolutional Neural Networks to detect Edge Localized Modes in DIII-D from Doppler Backscattering measurements

Predicting continuum breakdown with deep neural networks

Classification of the L-, H-mode, and plasma-free state: Convolutional neural networks and variational autoencoders on the edge reflectometer for KSTAR

A neural network for the analysis of Langmuir-probe characteristics

Neural Network Training for the Detection and Classification of Oceanic Mesoscale Eddies

Experimental identification of Ion Cyclotron Emission on HL-2A using YOLO neural network algorithm