NN-EFIT: Physics-constrained Plasma equilibrium reconstruction for magnetically confined Fusion�With the EFIT-AI Partnership
1
Scientific Achievement
Developed a physics-constrained neural network-based surrogate model to reconstruct plasma equilibrium (poloidal flux and toroidal current density) inside a Tokamak from experimental magnetic signals collected from DIII-D Fusion Facility.
Significance and Impact
Magnetic equilibrium is one of the most important information to understand the basic behavior of plasmas in magnetically confined plasmas. The developed neural network-based surrogate model serves as a reduced order model, thus opening avenues to accurate real-time plasma control.
Technical Approach
We developed a parallel python framework to extract the data from the DIII-D database to construct the I/O for NN-EFIT.
Developed a physics-informed data non-dimensionalization approach.
Proposed a hybrid architecture-penalty constraint approach in which poloidal flux is predicted by a neural network from magnetic signals and the force balance constraint is applied by another neural network using toroidal current density
NN-EFIT uses end-to-end neural architecture search to obtain uncertainty quantified predictions with accuracy and speed on par with real-time models and offline-EFIT models respectively.
Also generalized well to atypical discharges such as negative triangularity.
Fig.1: Reconstruction of both poloidal flux and toroidal current density
from magnetic signals
Lao, L.L., Kruger, S., Akcay, C., Balaprakash, P., Bechtel, T., Howell, E., Koo, J., Leddy, J., Leinhauser, M., Liu, Y. and Madireddy, S., 2022. Application of machine learning and artificial intelligence to extend EFIT equilibrium reconstruction. Plasma Physics and Controlled Fusion