Improved Surrogate Models for DIII-D Equilibrium Reconstruction and Tools for Uncertainty Analysis Delivered by EFIT-AI for the Theory and Simulation Performance Target

As part of the 2024 Theory and Simulation Performance Target (TSPT) the EFIT-AI project has made substantial improvements to surrogate models for DIII-D equilibrium reconstruction and tools for analyzing the uncertainty of fits to the experimental data. The accuracy of neural network (NN) models for plasma reconstructions has been improved by using larger training datasets and coupling analytic methods with advanced machine-learning techniques, such as a bi-level architecture and hyper parameter optimizations and quantified uncertainty from an ensemble of models. These advancements were enabled by an extension of the FAIR database of more than 6 million equilibria. In addition to the improved accuracy, the NN inference has been accelerated to achieve milli-second speed. These techniques have been used to develop surrogate models for kinetic equilibrium reconstructions, which include the poloidal flux, pressure, toroidal current density, safety factor, and other derived quantities such as beta and internal inductance. The capabilities of the Gaussian Process Regression (GPR) profile fitter with change-point kernel and the directive-based GPU accelerated EFIT solver have also been extended. Lastly, initial development of a standalone filament model for reconstructing 3D plasma responses is underway. Together, these improvements bring us closer to addressing equilibrium reconstruction challenges at DIII-D with methods that are transfereable to other tokamak experiments.