Directing TGLF saturation rule development with machine learning tools

The trapped gyro-Landau-fluid (TGLF) code solves a reduced set of gyrokinetic equations. To accurately model the nonlinear saturation of turbulence, the TGLF code employs so-called saturation rules SAT0, SAT1 and the recently developed SAT2. Using automated workflows within OMFIT, we have built a curated dataset of 25000 time and space slices of DIII-D plasmas that has been used to successfully validate these saturation rules. We find that SAT1 performs well when the heat flux is largely electrostatic, i.e. with an electromagnetic component satisfying Q_EM / Q_ES < 0.01. To help direct future saturation rule development, we have applied machine learning tools to our curated dataset. First, we applied model optimization tools to the free parameters in SAT0 and SAT1, which have been previously calibrated by the first-principles gyrokinetic code GYRO. We found that free parameters governing the importance of E × B shear could merit further attention during a possible recalibration. Second, we multiplied the wavenumber (k) spectrum of SAT1 with a hypothetical correction factor of the form a / k^c × exp( - b / k), where a, b, and c are the outputs of a neural network. After training this neural network to reproduce the experimentally inferred fluxes, we find that the amplitude of the correction factor is correlated with plasma parameters that are typically associated with trapped electron mode (TEM) turbulence. Therefore, future saturation model development could focus on TEM saturation.