17–22 May 2026
marinaforum REGENSBURG
Europe/Berlin timezone

2.091 Fine-Tuning of Machine Learning Diagnostics of Electron Density and Temperature from Helium Line Emissions

19 May 2026, 16:20
3h
Poster I. Plasma Edge and First Wall Diagnostics Postersession 2

Speaker

Shin Kajita

Description

While the helium line intensity ratio method has been used to measure electron, ne, and temperature, $T_e$, by combining measured line intensities with a collisional radiative model (CRM) [1], one of its difficulties is to include the photon transport and metastable atom transport. A machine learning (ML) approach has been considered as an alternative method to measure $n_e/T_e$ from the line ratios. If training data is sufficiently available, it can be a useful diagnostic tool. The challenging issue in this approach is developing a global model that can be applied to other devices. In this study, we collected an OES dataset and $n_e/T_e$ data from four linear divertor simulators, and we investigated the fine-tuning method to develop a global cross-machine model.
Data from the following four linear devices are used: Magnum-PSI, NAGDIS-II, and PISCES-A, and Lotus-I. Line emissions at 447.1, 492.2, 501.6 + 504.8, 667.8, 706.5, and 728.1 nm are used. The dataset includes 24, 64, 6, and 3 discharges (radial profiles) and 960, 417, 342, and 70 data points from Magnum-PSI, NAGDIS-II, PISCES-A, and Lotus-I, respectively. Laser Thomson scattering was used in Magnum-PSI and a Langmuir probe was used in the other devices to obtain $n_e/T_e$. In addition to a deep neural network (DNN) model, physics-informed ML approach [2] was also tested, where a pre-trained NN with a CRM tuned with experimental data [3].
It was shown that a DNN model trained with the dataset from three devices leads to an error of ~100%, when applying it to a remaining unseen fourth device for both ne and $T_e$, which is significantly higher than the model applied to the seen devices. This is primarily due to device-specific parameters such as plasma radius and different ranges of ne and $T_e$, which hinder the model’s generalizability across devices. To address this, we additionally performed fine-tuning using data from the target device itself. It was found that errors significantly decreased by the fine-tuning process with a small amount of data. Furthermore, the reduction in the errors was more significant for physics-informed models. The results suggested that the physics-informed model has an advantage when using fine-tuning with a limited dataset.

[1] S. Kajita, D. Nishijima, Journal of Physics D: Applied Physics 57 (2024), 423003.
[2] S. Kajita, https://arxiv.org/abs/2506.20117.
[3] M. Goto, J. Quantitative Spectroscopy and Radiative Transfer 76 (2003) 331.

Authors

Shin Kajita Daisuke NISHIJIMA (Center for Energy Research, University of California San Diego) Prof. Noriyasu Ohno (Nagoya University) Hirohiko Tanaka (Nagoya University) Dr Ivo Classen (DIFFER—Dutch Institute for Fundamental Energy Research, The Netherlands) Dr Quan Shi (The University of Tokyo) Dr Yuki Hayashi (The University of Tokyo) Keisuke Fujii (Fusion Energy Division, Oak Ridge National Laboratory, Oak Ridge, U.S.A) Motoshi Goto (National Institute for Fusion Science, Japan)

Presentation materials

There are no materials yet.