Development of a neural network-based model for electron temperature inference via modelled and experimental argon spectra measured by x-ray crystal spectrometer with an extension to tungsten spectra on EAST

Accurate inference of electron temperature (Te) profiles is essential for understanding energy transport and impurity dynamics in magnetic confinement fusion plasmas. Traditional spectral diagnostics based on satellite line ratios become unreliable in high- Te regimes due to reduced line intensity. Here we present a framework that combines Gaussian Process Regression (GPR) and convolutional neural networks (CNN) to reconstruct Te profiles from impurity emission spectra measured by a high-resolution x-ray crystal spectrometer on EAST. The GPR module leverages a non-stationary Gibbs kernel to interpolate spatially discrete ECE data with uncertainty quantification, while the CNN model learns the nonlinear mapping between impurity line intensity profiles and electron temperature. The model is trained and validated on a argon spectra dataset comprising over 2300 discharges and extended to tungsten spectra for high-Z impurity scenarios. It demonstrates robust performance across diverse plasma conditions, with the coefficient of determination (R2) > 0.9 and mean absolute errors below 0.5 keV. The model’s generalization capability is further confirmed using synthetic data generated by STRAHL simulations. Our results highlight the feasibility of real-time surrogate diagnostics for future fusion experiments and control systems, providing a scalable approach for high-temperature profile inference using x-ray spectrometer.