2024.05.20

　

The Plasma Quantum Process Unit holds seminars to disseminate the research activities of the Unit. This time, Mr. Ivan Traparić, University of Belgrade, Serbia, will give a lecture titled “Modeling of Tungsten EUV spectra with Variational Autoencoder and Tritium Retention Study using Laser Induced Breakdown Spectroscopy (LIBS) Technique”. We hope many of you will join us.

【Date】 Monday, June 10, 2024, 11:00-12:00

【Place】 Seminar Room (Room 501), 5th floor, Research I Building, National Institute for Fusion Science

【Zoom】 Registration at the following address is required to participate in this seminar via Zoom.
　https://us06web.zoom.us/meeting/register/tZMvcOiprDMiGtTkgLRx-GG79YlLtxGeKmBK

【Lecturer】 Mr. Ivan Traparić, University of Belgrade (Serbia)

【Title】 Modeling of Tungsten EUV spectra with Variational Autoencoder and Tritium Retention Study using Laser Induced Breakdown Spectroscopy (LIBS) Technique

【Abstract】

In the first part of the talk, the possibility of application of Variational Autoencoders (VAE) as a tool for generation and modeling of Extreme Ultraviolet (EUV) part of spectra of tungsten emitted from high – temperature plasmas was examined. Data used were the spectra recorded in the previous campaigns on Large Helical Device (LHD), where in total 1126 spectra in the range from 4 up to 7 nm were used. Plasma parameters that were considered for this investigation were central electron temperature, electron temperature at normalized radial position reff/a99 = 0.27, electron temperature at mid – radius, radiated power from plasma and time elapsed after pellet injection. Beta – Variational Autoencoder was used to map the spectra to the latent space with dimension of 3, while the optimal value of beta was 0.05. Experimental parameters were mapped to the latent space with Random Forest (RF) where coefficient of determination R2 = 0.9 and root mean square error (RMSE) RMSE = 0.07 were achieved. Additionally, to determine the electron temperature profile associated with the generated spectra, one – dimensional convolutional neural network (1D-CNN) was trained on original data. Used beta – VAE led to the R^2 reconstruction score of 0.91 and root mean squared error (RMSE) of 233 counts/ch. Developed model provides the possibility for quick generation of the expected spectra with only five input plasma parameters, gives opportunity for synthetic plasma diagnostics in terms of generation of electron temperature profiles, as well as foundation for further data augmentation and studies on the accumulation of impurities in fusion plasmas.

　　Second part of talk will be devoted to the application of LIBS technique to the study of tritium retention in the first wall of future fusion reactors, a project developed through the collaboration agreement between NIFS and University of Belgrade. As a start, based on available theoretical data, the condition for separation of tritium and deuterim Balmer alpha will be found. This condition will be further elaborated with respect to the specifications of spectrometer that are needed in order to experimentally resolve these two lines. Finally, the usage of TEA CO₂ laser to create plasma and finding of the optimal experimental plasma parameters will be explained.

【Contact】
Chihiro Suzuki (csuzuki-at-nifs.ac.jp)　*Please change “-at-” to @.