Neural network approximated Bayesian inference of edge electron density profiles at JET

Abstract

Abstract A neural network (NN) has been trained on the inference of the edge electron density profiles from measurements of the JET lithium beam emission spectroscopy (Li-BES) diagnostic. The novelty of the approach resides in the fact that the network has been trained to be a fast surrogate model of an existing Bayesian model of the diagnostic implemented within the Minerva framework. Previous work showed the very first application of this method to an x-ray imaging diagnostic at the W7-X experiment, and it was argued that the method was general enough that it may be applied to different physics systems. Here, we try to show that the claim made there is valid. What makes the approach general and versatile is the common definition of different models within the same framework. The network is tested on data measured during several different pulses and the predictions compared to the results obtained with the full model Bayesian inference. The NN analysis only requires tens of microseconds on a GPU compared to the tens of minutes long full inference. Finally, in relation to what was presented in the previous work, we demonstrate an improvement in the method of calculation of the network uncertainties, achieved by using a state-of-the-art deep learning technique based on a variational inference interpretation of the network training. The advantage of this calculation resides in the fact that it relies on fewer assumptions, and no extra computation time is required besides the conventional network evaluation time. This allows estimating the uncertainties also in real time applications.