Continuous-Time Bayesian Estimation of Tokamak Plasma States

A new machine learning algorithm has been developed to efficiently interpret plasma states from previous and current sensor data. This model addresses a limitation of existing data driven plasma profile prediction models, which rely on accurate sampling at fixed time intervals; in practice, such data is not readily available due to limitations on available diagnostics and noise. The proposed method uses a Bayesian framework to update model predictions at time steps when sensor data is available, and a neural network to continuously evolve predictions during intervals when information is unavailable. This model is more compatible with plasma data that is irregularly observed in both time and space, which enhances the accuracy of interpretation. This makes it a potentially useful tool for better control over plasma states in fusion devices.