Multicomponent Spectra Estimation with Deep Learning

A wide range of plasma diagnostics rely on the spectral analysis for inference of plasma properties and behavior. In optical emission spectroscopy the collected light is an integrated measurement along a line-of-sight and the resulting spectrum can be modeled by a sum of distinct components. When high RF power is applied to the plasma, like in Lower Hybrid Current Drive (LHCD), the emission spectrum of Deuterium and Hydrogen can be substantially modified by the presence of the RF electric field. That, in combination with the Zeeman effect due to the equilibrium magnetic field, Doppler and Stark broadening, results in a complex spectrum. Theoretical modeling of those effects can be computationally expensive, and the fitting can be challenging if the initial guess of the spectrum parameters is not close to the optimal solution. We present the use Deep Neural Networks (DNN) to estimate the physical parameters used to model the WEST spectra. The DNN estimation provides at least x100 computational speedup and can be used both as a direct estimation of the RF electric field or as an initial guess for a more refined model fitting.