Large database validation of TGLF-NN and multi-fidelity GKNN on NSTX, MAST-U and DIII-D plasmas

A comprehensive validation of the TGLF saturation rules has been carried out on a curated database of profiles from NSTX, MAST-U and DIII-D. The approach involves training a machine-learning surrogate of TGLF for each electromagnetic (EM) or electrostatic (ES) setting of SAT0, SAT1, SAT1geo, 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 (http://fuse.help). We gauge the accuracy of each TGLF saturation rule by evaluating the mean relative error (MRE) between the measured and predicted temperature and density profiles. Results show that the TGLF-NN SAT3 model is most accurate for NSTX L-modes (MRE = 18.3%) and H-modes (13.3%). The updated ExB shear model of SAT1geo, SAT2 and SAT3 shows improved profile prediction accuracy compared to the SAT1 and SAT0 models. This contrasts with results for DIII-D and MAST-U, where the most accurate models are SAT0 EM with the quench rule for H-modes and SAT1 EM for L-modes and negative triangularity. A multi-fidelity gyrokinetic neural network (GKNN) has been developed to address differences between the TGLF and QLGYRO models (e.g. when microtearing modes are present). GKNN is trained on a database of 35k QLGYRO fluxes generated with inputs from the DIII-D, NSTX and MAST-U tokamaks using 175k node hours on NERSC's perlmutter cluster over 3 years. Using the same validation approach described above we show that GKNN leads to a systematic improvement in profile prediction accuracy of SAT2 EM and SAT3 EM models for all plasma configurations. Specifically, this multi-fidelity approach identifies GKNN SAT2 as the most accurate model for NSTX L-modes (12.1%) and H-modes (10.6%), while GKNN SAT3 is the most accurate model for all configurations of DIII-D and MAST-U. This validates GKNN as a fast and accurate surrogate model of gyrokinetic turbulence.