"Neural Plasma Reconstruction from Diagnostic Imaging"
EPS Conference on Plasma Physics, June 2022
Abstract
Experiments like MAST collect vast amounts of diagnostic data, among these imaging data of the plasma and impurity emissions captured by a high speed camera system. These high temporal resolution images can be used for plasma monitoring and control, but are underutilised due to the difficulties in estimating relevant plasma parameters such as the shape and composition from such images. To that end, we leverage a neural network to estimate the 2D distributions of neutrals, electrons and temperature from images of plasma obtained from the high speed cameras inside the MAST vessel. Our networks infers the plasma parameters by learning the non-linear mapping between synthetic images of the plasma D-alpha emission and the distributions of the neutrals density, electron density and the electron temperature. These emissions are computed using the equation nn × ne × PEC (ne, Te) where PEC is the Photon Emissivity Coefficient obtained from ADAS. Our networks are composed of an image encoder net that takes the high speed images and encodes them to a latent vector. We also introduce an optional subnetwork to provide encoded latent vectors of point sample measurements that can be concatenated to the encoded image latent vectors with the goal of incorporating other diagnostic measurements and further improving accuracy. This latent vector is then decoded to predict the inputs of our plasma generation model. The biggest contribution of our networks are in predicting the distribution of neutrals in the vessel which cannot be measured directly and requires expensive Monte Carlo simulations to estimate.
Download
Bibtex
EPS Conference on Plasma Physics, June 2022
AbstractExperiments like MAST collect vast amounts of diagnostic data, among these imaging data of the plasma and impurity emissions captured by a high speed camera system. These high temporal resolution images can be used for plasma monitoring and control, but are underutilised due to the difficulties in estimating relevant plasma parameters such as the shape and composition from such images. To that end, we leverage a neural network to estimate the 2D distributions of neutrals, electrons and temperature from images of plasma obtained from the high speed cameras inside the MAST vessel. Our networks infers the plasma parameters by learning the non-linear mapping between synthetic images of the plasma D-alpha emission and the distributions of the neutrals density, electron density and the electron temperature. These emissions are computed using the equation nn × ne × PEC (ne, Te) where PEC is the Photon Emissivity Coefficient obtained from ADAS. Our networks are composed of an image encoder net that takes the high speed images and encodes them to a latent vector. We also introduce an optional subnetwork to provide encoded latent vectors of point sample measurements that can be concatenated to the encoded image latent vectors with the goal of incorporating other diagnostic measurements and further improving accuracy. This latent vector is then decoded to predict the inputs of our plasma generation model. The biggest contribution of our networks are in predicting the distribution of neutrals in the vessel which cannot be measured directly and requires expensive Monte Carlo simulations to estimate.
Download
- Ozturk_EPS2022.pdf (1.85Mb)
Bibtex
@conference{Ozturk:2022:NPR,
author = {Ozturk, Ekin and Akers, Rob and Ghosh, Abhijeet and Pamela, Stanislas and Peers, Pieter and , The MAST Team},
title = {Neural Plasma Reconstruction from Diagnostic Imaging},
month = {June},
year = {2022},
booktitle = {EPS Conference on Plasma Physics},
}
author = {Ozturk, Ekin and Akers, Rob and Ghosh, Abhijeet and Pamela, Stanislas and Peers, Pieter and , The MAST Team},
title = {Neural Plasma Reconstruction from Diagnostic Imaging},
month = {June},
year = {2022},
booktitle = {EPS Conference on Plasma Physics},
}