Augmenting time resolution of Thomson scattering profiles with machine learning in LTX-β

Thomson scattering (TS) is a key diagnostic tool for measuring spatially resolved electron density (n_e) and temperature (T_e) in all plasma confinement devices. However, due to complexity, TS is often difficult to implement for high time-resolution measurements, particularly on smaller machines. In the LTX-β tokamak, TS provides good spatial resolution but is limited to single time-point measurement per discharge, lacking any time resolution. To overcome this limitation, a neural network-based machine learning (ML) model is being developed to infer n_e and T_e spatial profiles at arbitrary time points within a discharge. Since n_e and T_e spatial profiles depend on various operational and diagnostic parameters, such as plasma current, shaping coil currents, fueling, and wall conditioning—these are used as model inputs. Notably, n_e measured by the microwave interferometer shows strong dependence on fueling rate and the extent of lithium wall conditioning in LTX-β. Therefore, line-integrated n_e from the interferometer is included as an input to encapsulate information about fueling and wall condition. The model is trained on a large, manually curated dataset spanning multiple years of LTX-β operation, consisting of TS data from 2000 plus discharges and at various time points and operational regimes. Results on model optimization and the accuracy of the predicted n_e and T_e profiles will be presented.