Transfer Learning and Ensemble Methods for Analyzing Thomson Scattering Spectra

Thomson scattering diagnostics are powerful methods to obtain measurements of electron temperature (T_e) and electron density (n_e). Current methods for analyzing Thomson scattering spectra, such as forward-fitting with analytical models, are computationally expensive, hindering real-time T_e and n_e measurements. We present a multilayer perceptron to calculate T_e and n_e from Thomson scattering spectra in the non-collective and collective regimes. The model uses a transfer learning technique by first training a base model with 10,000 synthetic spectra. Then, some hidden layers are retrained with experimental data. Our T_e and n_e measurements are the ensemble average of multiple model outputs. We take a second approach with a Bayesian neural network, which outputs probability distributions for T_e and n_e. We use these models to work towards real-time T_e and n_e measurements for high-repetition-rate experiments at the Phoenix Laser Laboratory and the Large Plasma Device.