Physics-assisted Machine Learning Spectral Fitting for Thomson Scattering Experiments on an Extreme Ultraviolet Plasma Light Source

Modern computer chips as well as the entire semiconductor industry rely on Extreme Ultraviolet Lithography (EUVL) at 13.5 nm ± 1% to create finer resolution features. In the industrial settings, the 13.5 nm photons are generated by a plasma following the interaction of 20-30 μm diameter molten tin droplets with focused CO₂ pulsed laser beams running at kHz repetition rates. Although the 13.5 nm light generation process has already been comprehensively studied numerically, only a handful of experimental studies report simultaneous measurements of the plasma parameters relevant to the production of the highly charged ions Sn⁸⁺‒Sn¹⁴⁺ responsible for the EUV light. Time-resolved collective Thomson scattering measurements, probing simultaneously the electron and ion features would provide a complete picture of the physics at play. To prepare experimental data analysis, a Matlab-based machine learning fitting tool was developed for real-time inference of the electron density, electron temperature and average charge state from Thomson scattering experimental spectra. “Neural Net Fitting”, an artificial neural network application from Matlab is deployed in conjunction with the analytical expression of the Thomson scattering spectral density function to perform a supervised non-linear regression model for the fitting of the experimental data.