Machine learning directed TGLF saturation rule development

To validate the TGLF model in general, and the saturation rules in particular, we have built a curated database of experimental profiles and power balance analyses of DIII-D plasmas using an automated workflow within OMFIT. After successful validation, we applied machine learning tools to this dataset to help direct future model development. As a first step, we applied model optimization tools to the free parameters in SAT0 and SAT1, which have previously been calibrated against a database of gyrokinetic simulations with GYRO. As a second step, machine learning tools were employed to direct SAT1 model development. The SAT1 intensity spectrum was multiplied with a `correction factor' of the form $(a/k^c)/\exp(b/k)$, where `k' is the binormal wavenumber, `b' is some constant and `a' and `c' are outputs of a neural network. Using system identification tools, we found that `a' and `c' are best described by a power law of plasma parameters that typically affect TEM turbulence. As third step, we used the above workflow independent of existing saturation rules to find new (or confirm existing) avenues of saturation rule development. A hypothetical saturation rule is constructed and all free parameters are calibrated against a database of gyrokinetic simulations. Comparison to experiment is then used to validate the development workflow of existing saturation rules.