Deep learning based plasma response models to 3D external magnetic field perturbations in EAST
Open Access
- 20 May 2026
- journal article
- Published by IOP Publishing in Nuclear Fusion
- Vol. 66 (7)
Abstract
The real-time response monitoring model and the physics feature based response optimization model are proposed on EAST using three deep learning models. Compared with the widely used fully connected neural network (FNN) and convolutional neural network -- long short-term memory (CNN--LSTM) models, the inverted transformer (iTransformer) model, which uses the attention mechanism to more naturally capture multivariate dependencies, demonstrates the best overall performance. For real-time response monitoring, the iTransformer model is trained on 133157 experimental samples with 4-fold toroidal phase symmetry augmentation. This model establishes a direct nonlinear mapping from resonant magnetic perturbation (RMP) coils and electromagnetic measurement sensors to the plasma response. With an inference time per discharge of less than $10~\mathrm{ms}$, it can provide real-time response monitoring under flexible 3D fields. For experimental response optimization, the physics feature based iTransformer model is trained on 12080 filtered experimental samples. This model captures the dependence of the response amplitude on key parameters, including $I_{\mathrm{RMP}}$, $\phi_{\mathrm{UL}}$, $q_{95}$, $\beta_{N}$, $n_{e}$, $B_{t}$, and $l_{i}$. Combined with SHapley Additive exPlanations (SHAP) analysis, it provides interpretable insights into parameter influences and directly guides the optimization of 3D field experimental discharges. This work establishes a comprehensive framework for real-time monitoring and physics feature based optimization of plasma responses based on statistical patterns in large datasets, thereby paving the way for active control of error-field locked mode and edge localized mode (ELM).
Keywords
Funding Information
- National Natural Science Foundation of China (12435014, 12575248, 12175276)
- The National MCF Energy R&D Program of China (2024YFE03010000, 2024YFE03270800)
- the HFIPS Director's Fund (BJPY2022B05)
- The Strategic Priority Research Program of Chinese Academy of Sciences (XDB0790100)
This publication has 61 references indexed in Scilit:
- Exploring parameter dependencies in fusion experiments using interpretable machine learning: A case study on core electron temperature in EASTFusion Engineering and Design, 2025
- Multimodal super-resolution: discovering hidden physics and its application to fusion plasmasNature Communications, 2025
- MHD spectrogram contribution to disruption prediction using Convolutional Neural NetworksFusion Engineering and Design, 2024
- Highest fusion performance without harmful edge energy bursts in tokamakNature Communications, 2024
- Disruption prediction and analysis through multimodal deep learning in KSTARFusion Engineering and Design, 2024
- Nonlinear evolution and control of neo-classical tearing mode in reversed magnetic shear tokamak plasmasReviews of Modern Plasma Physics, 2022
- Magnetic control of tokamak plasmas through deep reinforcement learningNature, 2022
- Predicting disruptive instabilities in controlled fusion plasmas through deep learningNature, 2019
- Calibration and uncertainty analysis of magnetic measurement for plasma shape reconstruction on EASTFusion Engineering and Design, 2016
- Advantages and disadvantages of using artificial neural networks versus logistic regression for predicting medical outcomesJournal of Clinical Epidemiology, 1996