Exploring the potential of machine learning for real-time neutron emissivity tomography using the Vertical Neutron Camera of ITER
•ML algorithms enable real-time neutron emissivity tomography.•Linear regression and XGB meet 10 % accuracy and 1 ms resolution requirements of ITER.•XGB model preferred for real-time neutron yield rate and fusion power analysis.•NN accurately reconstruct neutron emissivity profiles in high-yield sc...
Gespeichert in:
Veröffentlicht in: | Fusion engineering and design 2024-07, Vol.204, p.114519, Article 114519 |
---|---|
Hauptverfasser: | , , , |
Format: | Artikel |
Sprache: | eng |
Schlagworte: | |
Online-Zugang: | Volltext |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Zusammenfassung: | •ML algorithms enable real-time neutron emissivity tomography.•Linear regression and XGB meet 10 % accuracy and 1 ms resolution requirements of ITER.•XGB model preferred for real-time neutron yield rate and fusion power analysis.•NN accurately reconstruct neutron emissivity profiles in high-yield scenarios.•Study suggests ML as a viable solution for ITER diagnostics.
The Vertical Neutron Camera (VNC) is one of the key neutron diagnostic systems of the ITER tokamak designed to measure and reconstruct the time- and space-time-resolved neutron emissivity poloidal profile. The ability to obtain the properties of a neutron emissivity in real-time will help not only to study the processes occurring in plasma during discharge, but also to control them.
The main problem of existing plasma tomography methods (e.g. maximum likelihood estimation, Tikhonov regularization) is their computational complexity, which significantly hinders the effort to meet the time resolution and latency requirements set for neutron diagnostics of the ITER tokamak. This motivated us to investigate machine learning algorithms as a potential solution for real-time tomography, wherein the key distinction lies in the separation of the fitting and prediction processes. Once fitted, these models have the capability to infer hundreds of objects per second, thereby paving the way for their real-time application.
In this paper, we have demonstrated the possibility of using machine learning methods to reconstruct the time-resolved (neutron yield rate and fusion power) and space-time-resolved (fusion power density and neutron emissivity profile) characteristics of the neutron source with real-time operation capabilities. Linear regression and extreme gradient boosting methods reconstructed the DD-neutron yield rate, the DT-neutron yield rate and fusion power with a mean absolute percentage error (MAPE) of less than 10 % in less than 1 ms. Deconvolutional neural network reconstructed DD- and DT-neutron emissivity profiles in ITER scenarios with the phantom neutron plasma sources with MAPE averaged over the VNC field of view of less than 10 % in approximately 5 ms. |
---|---|
ISSN: | 0920-3796 1873-7196 |
DOI: | 10.1016/j.fusengdes.2024.114519 |