Speaker
Marco Cecconello
(Uppsala University)
Description
The ITER Radial Neutron Camera (RNC) is a diagnostic with multiple collimated inputs aiming at characterizing the neutron source. The RNC plays a primary role in the advanced control measurements and physics studies of ITER, and acts as backup for system machine protection and basic control measurements. The RNC primary design driver is the measurement of the neutron emissivity radial profile within specified measurement requirements regarding temporal and spatial resolution and fusion power.
This paper presents a method based on neural network methods to provide an estimate of the neutron emissivity profile in different deuterium-tritium ITER scenarios and for different RNC architectural configurations which are under investigation [1]. The design and optimization of the feed-forward neural network with back-propagation algorithmand the choice of the training data sets will be discussed. The effect of statistical noise and background are included in the neural network supervised learning phase. A decision algorithm has been implemented to select which inverted neutron emissivity profile gives the best estimate of the real one.
The profile recognition based on neural networks is sufficiently fast that it is considered feasible for a real time environment [2]. This study indicates that neural networks can achieve an accuracy and precision within the spatial and temporal requirements set by ITER. The following aspects of the neural network implementation will be discussed: i) the decision algorithm requirements of a priori knowledge of the plasma flux surfaces; ii) the role of ensemble averaging of multiple static predictors and iii) the effect of missing data.
[1] D. Marocco et al., System Level Design and Performances of the ITER Radial Neutron Camera, IAEA 2016
[2] N. Cruz et al., The Real-Time Software Design for the ITER Radial Neutron Camera, this conference.
Co-author
Marco Cecconello
(Uppsala University, Uppsala, Sweden)
