Model Order Reduction of the Plasma Equilibrium Reconstruction framework EFIT with Deep Neural Networks

We present a model order reduction (MOR) of magnetics-only EFITs with neural network (EFIT-MORNN) surrogates that have been trained on the 2019 DIII-D data. Our neural networks reconstruct DIII-D equilibria, given an input space comprising external measurements of the poloidal magnetic field, poloidal flux, plasma current, and currents in the external poloidal-field coils. A total of 160 thousand magnetic equilibria from 2019 were used. First, we built a physics-informed machine-learning (ML) surrogate based on a combination of a convolutional neural network and a multi-layer perceptron (MLP) used to concurrently predict the equilibrium poloidal magnetic flux y and the toroidal current density J_f. Next, two additional MLP-based surrogate models were built and trained to reconstruct the plasma boundary shape and global quantities including the normalized beta, the internal inductance, and the edge safety factor. These trained networks were then used to infer plasma equilibria for 4 different types of DIII-D discharges: high poloidal beta, hybrid, super H-mode, and negative triangularity. Performance of EFIT-MORNN was compared to offline-EFIT as well as real time (RT) EFIT. The trained EFIT-MORNN reconstructed y with better than 99% accuracy and J_f better than 98% accuracy, showing an improvement over RT-EFIT reconstructions, with a reconstruction speed per time slice that is at least comparable to that of RT-EFIT. The offline-EFIT-like accuracy and RT-EFIT-like speed offer the possibility of turning EFIT-MORNN into a real-time tool. We also present a new framework that initializes the EFIT iteration loop with an equilibrium generated by EFIT-MORNN to increase the accuracy of RT-EFITs.