Fast and Accurate Diagnosis of Electron Temperature by a Machine Learning Model Trained with Time-Evolved Plasma Spectra

Machine learning has demonstrated as a powerful tool for rapid plasma diagnostics but is often challenged by low prediction accuracy attributed to an inadequate volume of training samples. This work overcomes this limitation by introducing a straightforward method to efficiently generate extensive training samples from time-resolved plasma spectra. By linearly interpolating spectral line intensities between adjacent delay times and applying the Saha–Boltzmann method, the training data featuring diverse temperature labels are generated. The characteristics of interpolated intensities and temperature labels were further analyzed and the results demonstrate similar characteristics to the experimentally measured data. An artificial neural network model was successfully trained by this dataset and the trained model was tested by leave-one-out method. The test result confirms that the model achieves very high prediction accuracy (0.18% relative error) and the trained model possesses the ability to identify the electron temperature beyond the range covered by the training dataset. This study significantly contributes to the rapid and accurate diagnosis of plasma temperature, facilitating the development of industrial online optimization in applications like pulsed laser deposition and arc welding.