Transfer Learning Analysis of Collective and Non-Collective Thomson Scattering Spectra

Thomson scattering (TS) diagnostics provide reliable, minimally perturbative measurements of fundamental plasma parameters, such as electron density ($n_e$) and electron temperature ($T_e$). Deep neural networks (DNNs) can provide accurate $n_e$ and $T_e$ estimates from when conventional fitting algorithms may fail, such as when TS spectra have high signal noise, or when fast analysis is required for real-time diagnostics. A drawback of DNNs is that they typically require large training sets. However, a DNN trained on synthetic TS spectra can be adapted to analyze experimentally measured TS spectra via transfer learning. Here, we present five distinct DNNs trained with transfer learning to estimate $n_e$ and $T_e$ in both the collective and non-collective scattering regimes. The synthetic non-collective spectra are generated from a Gaussian model, with the relationship between total signal intensity and $n_e$ determined by Raman scattering calibration. The synthetic collective spectra are generated from PlasmaPy's spectral density function. We quantify the appropriate use case of transfer learning by comparing the error in $n_e$ and $T_e$ estimates between models trained with and without transfer learning, and we observe improvement when the training set contains less than approximately 200 experimentally measured spectra.