Impact of Various DIII-D Diagnostics on the Accuracy of Neural Network Surrogates for Kinetic EFIT Reconstructions*

Equilibrium reconstructions compute key plasma physics parameters such as global scalar variables and the flux contours, amongst other quantities. Kinetic equilibrium reconstructions make use of additional diagnostic data internal to the plasma to compute key internal profiles such as the pressure and current. Altogether, these quantities form a self-consistent equilibrium reconstruction and are crucial for further downstream analysis of plasma phenomenon, transport codes, and feedback into actuator control. Present-day kinetic equilibrium reconstructions are time consuming and therefore surrogate models are being developed to produce faster reconstructions while maintaining a high accuracy. This work develops a multi-layer perceptron (MLP) neural network (NN) model framework as a surrogate for kinetic Equilibrium Fitting (EFITs). It trains on a curated database of automatically generated kinetic EFITs from the 2019 DIII-D discharge campaign. This work investigates the impact of including different categories of diagnostic data and machine actuator controls as input into the NN and their impact on the accuracy of the inference. For all the models (MLP NNs with different configurations) in this work, the predictions on multiple different EFIT solutions such as the poloidal magnetic flux, global scalar quantities, pressure and current profiles, are highly accurate. Looking at different permutations of categories of diagnostics data reveals interesting drivers of accurate predictions. Namely, the magnetics-only model infers accurate kinetic profiles and the inclusion of additional data does not significantly impact the prediction accuracy. When the models are re-designed and tasked with inferring only a single EFIT solution such as the pressure profile or current profile, there is a noticeable increase in accuracy of the prediction as more data is included. These results indicate that certain MLP NN configurations can be reasonably robust to different burning-plasma-relevant diagnostics depending on the individual accuracy requirements for the equilibrium reconstruction tasks.