Real-time measurement and feedback control of key plasma parameters are critical for future fusion reactor operation, with ion temperature being a vital control target as part of the triple product for fusion ignition. However, plasma diagnostics tends to require complex data analysis. To acquire ion temperature T_i from charge exchange recombination spectroscopy (CXRS), a widely used method is through iterative spectral fitting, which is time-consuming and calls for expert intervention during data analysis. On top of that, frequent human expert intervention is needed in the conventional iterative fitting. Therefore, the conventional method cannot meet the meet the demand for real-time T_i measurement. Neural Networks (NN), which is capable of learning the underlying relationships between the measured spectra and T_i, is a promising approach to cope with this problem. In fact, NN approaches have been widely adopted in the field of magnetic confined plasma. Previous study in JET has achieved a satisfactory accuracy for inferring T_i from CXRS spectra compared to the conventional fitting results. Recently the study of disruption prediction has achieved great progress with the help of deep neural networks. However, these researches are conducted in steadily-operating devices, where for NN models, the data distribution is similar in training set and test set. This is not the case for newly-built tokamak like HL-3, or for future fusion reactors such as ITER. For new devices, there will be a period for the plasma parameters to raise from low to high ranges. In this case, investigating the extrapolation capability of NN models based on low parameter training data is of paramount importance.
A Convolutional Neural Network (CNN)-based model is proposed to accelerate the analysis of spectral data of CXRS, with a focus on investigating the model’s extrapolation capability to much higher T_i ranges. The dataset consists of about 122 thousand pieces of spectral data, along with their corresponding inferred T_i from offline iterative process. The results demonstrate that the CNN-based model provides excellent T_i analysis and reduces the inference time for analyzing a single spectrum to less than 1 ms, which is 100-1000 times faster compared to traditional spectral fitting method. However, the performance of the data-driven neural network model is limited by challenges such as insufficient data and imbalanced data distribution, which further deteriorates the extrapolation capability. Generally, data with higher T_i constitute a small percentage of the total dataset. In the case of our study, only about 5% of the spectra correspond to T_i > 2 keV (among 2-4 keV). Yet they reflect the temperature of central plasma, which is more important for assessing the performance of plasma. To overcome this limitation, the study synthesizes high-temperature data based on experimental data from discharges with T_i in low-temperature range. By incorporating 5% synthetic data into the training set only consisting of data with T_i<2 keV, the model’s extrapolation capability is extended to cover the whole range of T_i < 4 keV. The mean relative error of the mode in 3 keV < T_i < 4 keV is reduced from 35% to below 15%. This approach demonstrates the feasibility of using synthetic data to enhance the performance of artificial intelligence algorithms in the field of magnetic confinement fusion. The findings provide valuable insights for the development of real-time ion temperature measurement and feedback control for future high-parameter fusion devices. Furthermore, the study lays a foundation for research in areas that require high-performance across-device characteristic, such as machine learning-based disruption prediction and tearing mode control.