1 of 1

Physics-informed neural surrogate with Green’s function augmentation for consistent plasma equilibrium reconstruction.�With the EFIT-AI Partnership (FES)

Scientific Achievement

Developed a method for physics-informed neural network surrogate model for the plasma current and poloidal magnetic flux in a tokamak fusion device exploiting the separation of the total poloidal flux into a known external contribution due to the poloidal field coils & unknown internal contribution due to the plasma current.

Significance and Impact

Including Green’s function tables in the prediction physically enforces consistency between toroidal current density and poloidal flux and the synthetic diagnostics. This allows for generalization of our physics-informed neural network prediction well outside the training set with difference set of magnetic diagnostics, such as from DIII-D (R=1.7 m) to ITER (R=6.2 m).

Technical Approach

Represent the plasma current as a linear combination of basis functions using PCA of plasma toroidal current densities of the DIII-D EFIT-AI equilibrium.
Utilizing EFIT's Green's function tables, basis functions are created for the poloidal flux and diagnostics generated from the toroidal current.
A probabilistic neural architecture search (NAS) used to train a neural network to learn 32 coefficients, significantly simplifying the inference problem.

Fig.2: Reconstruction of both poloidal flux and toroidal current density for ITER trained on DIII-D toroidal current density

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): J. McClenaghan, C. Akcay, T. B. Amara, X. Sun, S. Madireddy,�L. L. Lao, S. Kruger, O. M. Meneghini, and the EFIT-AI team., 2024 Augmenting Machine Learning of Grad-Shafranov Equilibrium Reconstruction with Green’s Functions. To be Submitted, Physics of Plasmas (PoP)

Fig.2: Example contours DIII-D equilibria and ITER equilibria

J. McClenaghan, C. Akcay, T. B. Amara, X. Sun, S. Madireddy, L. L. Lao, S. Kruger, O. M. Meneghini, and the EFIT-AI team., 2024. “Augmenting Machine Learning of Grad-Shafranov Equilibrium Reconstruction with Green’s Functions”. To be Submitted, Physics of Plasmas (PoP)

LOCAL POC: Sandeep Madireddy (ANL)

TALKING POINTS:

We present a model order reduction of magnetics-only EFITs with neural network surrogates predicting 32 basis function rather than 129x129 poloidal flux grid points (see Jt equation on right).
This is done by taking advantage of Green’s function tables that EFIT uses. Split poloidal flux into known external coil contribution computed by Green’s functions (psi equation on the right). Unknown plasma contribution of poloidal flux is written as function of same basis functions as toroidal current (psi_pla equation on right).
Additionally, synthetic diagnostics such as magnetic probes and poloidal flux loops can be written as as function of same basis functions as toroidal current (Cm equation)
NAS trained to learn 32 coefficients, significantly simplifying the inference problem. This showed similar predictive capability as more complication architecture for DIII-D magnetics database.
This method generalizes extremely well, where you can predict yet to be built tokamak ITER equilibrium only using input data from DIII-D. ITER is much larger tokamak with completely different set of diagnostic inputs.
To predict ITER, we take the gridded DIII-D toroidal current densities and apply the ITER Green’s function tables to create a training database of equilibria. Training NAS on this database, we can predict ITER equilibria Equilibrium prediction on the right.

METADATA:

Name of the associated awarded project: EFIT-AI, SciDAC Rapids2

PI name(s): Lang Lao(GA), Sandeep Madireddy(ANL), Scott Kruger (TechX), Sam Williams (LBNL)

Name of the program manager: Kalyan Perumalla

CITATIONS:

J. McClenaghan, C. Akcay, T. B. Amara, X. Sun, S. Madireddy, L. L. Lao, S. Kruger, O. M. Meneghini, and the EFIT-AI team., 2024 Augmenting Machine Learning of Grad-Shafranov Equilibrium Reconstruction with Green’s Functions. To be Submitted, Physics of Plasmas (PoP)

AWARDS:

N/A

REPRODUCIBILITY:

Repo not made public yet

BACKGROUND: