Large database validation of TGLF on DIII-D and MAST-U plasmas

A comprehensive validation of the TGLF saturation rules has been carried out on a curated database of 7500 profiles from DIII-D and 2000 profiles from MAST-U. The approach involves training a machine-learning surrogate of TGLF for each tokamak and for each electromagnetic (EM) or electrostatic (ES) setting of SAT0, SAT1, SAT2 and SAT3. The resulting TGLF-NN models are an ensemble of 20 neural networks trained on a database of 10 million TGLF simulations and are 1000 times faster than the original TGLF model at predicting the temperature, density and rotation profiles in the database. These predictions are carried out in a flux-matching transport solver within the FUSE integrated modeling suite. We evaluate the accuracy of each TGLF saturation rule by evaluating the mean absolute percentage error (MAPE) between the stored energy, central temperatures and densities calculated from the experimental and predicted profiles. Results show that the SAT1 EM model is most accurate for DIII-D L-modes (MAPE = 22%) and negative triangularity (15%), while SAT0 EM with the quench rule is most accurate for H-modes on DIII-D (15%) and MAST-U (19%), while MAST-U L-modes are best described by SAT0 ES (26%). Finally, a multi-fidelity neural network model of gyrokinetic turbulence (GKNN-0) has been developed to address differences between the linear TGLF and CGYRO models (e.g. when microtearing modes are present). GKNN-0 is trained on a database of 12k linear CGYRO spectra generated with inputs from the DIII-D and MAST-U tokamaks using 50k node hours on NERSC's perlmutter cluster. Using the same validation approach described above we show that GKNN-0 leads to a systematic improvement in profile prediction accuracy of SAT2 EM and SAT3 EM models for all plasma configurations. While the largest improvements are seen for MAST-U plasmas, this multi-fidelity approach also allows SAT2 EM to replace SAT1 EM as the most accurate model for DIII-D L-modes (21%) and negative triangularity (14%). This validates GKNN-0 as a fast and accurate surrogate model of gyrokinetic turbulence.