1 of 1

EFIT-PRIME: Probabilistic and physics-constrained reduced-order neural network model for equilibrium reconstruction in DIII-D �With the EFIT-AI Partnership (FES)

1

Scientific Achievement

Developed a probabilistic surrogate model that employs neural architecture search-based uncertainty quantification and integrates physics constraints from the Grad-Shafranov equation, improving prediction reliability and showing high generalizability by accurately forecasting extreme plasma shapes unseen in training.

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 EFIT-Prime model serves as a reduced order model, thus opening avenues to reliable and accurate real-time plasma control and analysis in fusion pilot plant .

Technical Approach

  • We developed a EFIT-AI database of high-fidelity DIII-D Equilibria.
  • Proposed a multi-model 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
  • EFIT-Prime employs probabilistic neural architecture search to obtain uncertainty quantified predictions with a separation of aleatory (data) and epistemic (model) uncertainty to interpret and explain the model predictions and confidence.

PI(s)/Facility Lead(s): Lang Lao (GA)

Collaborating Institutions: Argonne National Lab, General Atomics, Tech-X

ASCR Program: SciDAC Rapids2

ASCR PM: Ceren Susut

Publication(s): S. Madireddy, C. Akçay, S. E. Kruger, T. B. Amara, X. Sun, J. McClenaghan, J. Koo,5 A. Samaddar, Yueqiang Liu, Prasanna Balaprakash, and L.L. Lao,., 2024 EFIT-PRIME: Probabilistic and physics-constrained reduced-order neural network model for equilibrium reconstruction in DIII-D. To be Submitted, Physics of Plasmas (PoP)

S. Madireddy, C. Akçay, S. E. Kruger, T. B. Amara, X. Sun, J. McClenaghan, J. Koo,5 A. Samaddar, Yueqiang Liu, Prasanna Balaprakash, and L.L. Lao,., 2024 EFIT-PRIME: Probabilistic and physics-constrained reduced-order neural network model for equilibrium reconstruction in DIII-D. To be Submitted, Physics of Plasmas (PoP)

Fig.1: Reconstruction of both poloidal flux and toroidal current density

from magnetic signals

​