AI-based prediction and control of tokamaks: combining simulations and experimental data

Abstract

A unified AI (artificial intelligence) approach to predict and control the dynamics of kinetic plasma profiles in fusion reactors is presented. On one hand, it is demonstrated that empirical models trained on experimental data ("data-driven models") significantly outperform the state-of-the-art ASTRA and TRANSP codes ("simulations") when predicting within the distribution of the training set. On the other hand, it is demonstrated that simulations can perform as well or better than data-driven models when extrapolating outside of the training distribution. Multiple AI-based methodologies for combining the data-driven models and simulations, leveraging data from multiple machines (DIII-D and AUG), are presented. One of the methodologies better extrapolates to new regimes than either data-driven models or simulations alone. Applications of the holistic approach to the task of commissioning a new reactor such as ITER are discussed. A successful model-predictive control test at DIII-D based on the methodology is described.