Overview and Progress of the EFIT-AI Project

The EFIT-AI project is creating a modern advanced equilibrium reconstruction code suitable for tokamak experiments and burning plasmas. Key elements of the EFIT-AI framework include: 1. improved optimization and data analysis capabilities using a ML-enhanced Bayesian framework, 2. a Model-Order-Reduction (MOR) version of the two-dimensional (2D) Grad- Shafranov equation solver using a neural network, and 3. a MOR version of three-dimensional (3D) perturbed equilibrium reconstruction tool. The success of ML relies on large amounts of quality data. To enable this, we have created a large database of curated EFIT equilibria, which has required developing new curation techniques. We will present results of training EFIT-MOR on both magnetic and MSE reconstructions and discuss the applicability of EFIT-MOR for RT control and as initial conditions for EFIT-AI. The core solver, now called EFIT-AI, has been substantially improved in portability, performance, and testing, and is ready for use in production environments. Bayesian approaches using Gaussian Process techniques offer a way of accurately and efficiently providing the experimental inputs into EFIT-AI without the manual efforts required for removing bad data or avoiding overfitting. We include work on using new kernels and likelihood functions to improve accuracy of prior methods used within the fusion community.