19–21 Oct 2026
Frascati
UTC timezone
The deadline for the abstracts submission has been delayed to July 20, 2026

A multi-window neural-network model for line-profile-based temperature estimation in low-temperature plasma optical emission diagnostics

19 Oct 2026, 16:10
1h
Corridor (Frascati)

Corridor

Frascati

Poster AI in support of diagnostics Poster session A

Speaker

Natalja Zorina (Institute of Atomic Physics and Spectroscopy, Faculty of Science and Technology, University of Latvia)

Description

Optical emission spectroscopy is widely used for plasma diagnostics, but profile-based temperature estimation remains sensitive to line selection, instrumental broadening [1], noise, baseline distortions and overlapping spectral features. In this work, a multi-window artificial neural-network architecture is developed for spectrum-level temperature estimation from emission spectra of high-frequency electrodeless discharge lamps (HFEDL). The target temperature is defined as a gas-temperature-related line-profile parameter associated with the Doppler broadening of emission lines.
The model was trained, validated and tested on a controlled synthetic dataset with known ground-truth temperature values. Its performance was evaluated using MAE, RMSE, R², exact temperature-class accuracy and neighboring-class accuracy. The proposed multi-window approach was compared with single-window CNN, full-spectrum resampled CNN and feature-based Random Forest/Extra Trees baselines. The attention-based variant also provides information on the relative contribution of individual spectral windows, supporting model interpretability in line with recent developments in machine-learning-assisted optical emission diagnostics [2].
The developed software prototype integrates spectral-window extraction, neural-network training, validation, testing, inference and visualization. The results demonstrate the feasibility of a TRL4-level AI-supported diagnostic model for line-profile-based temperature estimation under controlled laboratory conditions, providing a basis for further comparison with classical profile-fitting methods and experimental HFEDL spectra.

The work was supported by 1.1.1.9 Research application No 1.1.1.9/LZP/1/24/023 of the Activity “Post-doctoral Research” “Developing an artificial neural network model to analyze emission spectra of high-frequency electrodeless lamps.”

[1] Zorina N. (2010). Deconvolution of the spectral line profiles for the plasma temperature estimation, Nuclear Instruments and Methods in Physics Research A, 623, 763–765
[2] Kajita S. et al.,(2023) Application of machine learning for optical emission spectroscopy data in NAGDIS-II, Fusion Engineering and Design, 196,114012.

Author

Natalja Zorina (Institute of Atomic Physics and Spectroscopy, Faculty of Science and Technology, University of Latvia)

Co-author

Gita Revalde (1)Institute of Atomic Physics and Spectroscopy, Faculty of Science and Technology, University of Latvia2)Institute of Physics and Materials Science, Faculty of Natural Sciences and Technology, Riga Technical University)

Presentation materials