Self-Attention-based Sequential Plasma Evolution Simulator for EXL-50U Tokamak

Modeling the plasma evolution for a tokamak device is difficult, and requires numerous assumptions and linearization based on a conventional physical principled model. Recently, learning-based approaches are promising in plasma evolution prediction. However, existing models usually employ LSTM and CNN1d to predict limited plasma parameters, which may not effectively extract temporal information from sequential inputs. To this end, we propose a mult-head self-attention neuron network for plasma evolution in this paper. Firstly, we propose a new framework for plasma evolution simulation that integrates deep learning modules and physic-principled plasma equilibrium fitting methods to achieve plasma parameters and shape evolution. Secondly, we develop a new multi-head self-attention-based model for plasma evolution prediction, which can better capture temporal connections and focus on relevant inputs adaptively than LSTM and CNN1d. Finally, we validate our proposed model in a real-world tokamak device named EXL-50U, and compare the proposed model with different existing learning-based models. The experimental results show that the proposed SSPES can obtain lower accumulated prediction errors in all measurements, e.g., about 70% improvement in plasma current and 17% improvement in poloidal magnetic flux from LSTM.