GS-DeepNet: mastering tokamak plasma equilibria with deep neural networks and the Grad-Shafranov equation

Area of Science:

• Plasma physics
• Fusion energy research
• Computational science

Background:

• Achieving a sustained burning-plasma state in magnetically confined fusion reactors requires maintaining plasma equilibrium.
• Plasma equilibrium, a force balance between Lorentz and pressure gradient forces, is crucial for fusion energy.
• Current methods for identifying plasma equilibria are computationally intensive and involve subjective human decisions.

Purpose of the Study:

• To develop a novel, unsupervised learning method for real-time identification of plasma equilibria.
• To overcome the limitations of traditional iterative numerical approaches in plasma equilibrium reconstruction.
• To enable more reliable and precise control of fusion-grade plasmas.

Main Methods:

• Introduction of GS-DeepNet, a dual-neural network architecture for unsupervised learning of plasma equilibria.
• One neural network generates equilibrium candidates adhering to Maxwell's equations.
• A second neural network enforces the force balance condition, guided by magnetic measurements.
• The system learns without traditional numerical algorithms or human subjective input.

Main Results:

• GS-DeepNet successfully learns and identifies plasma equilibria solely through unsupervised learning.
• The method provides reliable equilibria with quantifiable uncertainties, unlike conventional techniques.
• Demonstrated ability to learn without iterative numerical solvers or subjective exclusion of data.

Conclusions:

• GS-DeepNet offers a significant advancement in real-time plasma equilibrium identification for fusion energy.
• The unsupervised learning approach enhances accuracy and reduces reliance on human intervention.
• This method holds promise for improved control and optimization of fusion reactors, including ITER.