Study of line spectra emitted by hydrogen isotopes in tokamaks through Deep-Learning algorithms

• A Deep-Learning model is applied to Balmer-alpha spectra emitted by a hydrogen–deuterium plasma in tokamaks. • The applied model is a One-dimensional Convolutional Neural Network. • The input data is a set of theoretical spectra generated for parameters varying in large domains covering conditions of tokamak divertor plasmas. • The presented work is a proof-of-principle of the use of 1D-CNN to raw spectra for predictions (the isotopic ratio here). Artificial Intelligence (AI) is increasingly used in various plasma physics topics, including applications in spectroscopy and diagnostics in magnetically confined fusion plasmas. The paper focuses on the application of the convolutional neural network (CNN) algorithm to emission spectroscopy from the divertor regions of magnetic fusion devices. Specifically, we use a CNN to determine hydrogen isotopic ratios from the theoretical emission spectra of the Balmer-α line in hydrogen–deuterium (HD) plasmas. The motivation for coupling AI with spectroscopy is to develop novel frameworks that can outperform existing classical methods based on spectral line fitting, in terms of accuracy, speed, or adaptability. In a previous work, we have used a fully connected neural network algorithm for theoretical Hα/Dα line spectra emitted by HD plasmas which have been generated for conditions relevant to divertor plasmas in tokamaks (magnetic field, neutral temperatures and fractions and hydrogen concentration). The generated spectra were preprocessed to extract few spectroscopic features which were then used as input data by the neural network. In the present work, we apply for the first time a CNN model to raw synthetic Hα/Dα line spectra theoretically emitted by HD plasmas to predict the corresponding isotopic ratios. In this context, we show that the trained CNN predicts hydrogen isotopic ratios with deviations of up to 5% from the true values. Additionally, our model can generalize its predictions to spectra corresponding to any observation angle relative to the magnetic field, despite being trained solely on spectra from parallel observations. The prediction accuracy in these cases is comparable to the training accuracy.