Multi-fidelity neural network representation of gyrokinetic turbulence*

This presentation will introduce a multi-fidelity neural network model of gyrokinetic turbulence GKNN-0, which has been trained and validated against a database of 5 million TGLF simulations and 5000 linear CGYRO simulations with experimental input parameters from the DIII-D tokamak. The first half of the presentation will review the TGLF saturation rules - SAT0, SAT1, SAT2 - and present a big data approach to validating both the linear model of TGLF and the saturation rules using experimental data from the DIII-D and MAST-U tokamaks. As a highlight, the TGLF model is shown to accurately reproduce both experimental and quasi-linear CGYRO fluxes for electrostatic turbulence, and the SAT2 model shows improved accuracy compared to SAT1 at capturing physics of trapped electron modes and the instability threshold of kinetic ballooning modes. The second half of the presentation will focus on database generation and benchmarking of a single-fidelity, retrained TGLF-NN and a multi-fidelity GKNN-0. The training database uses synthetically extended data from DIII-D as input, and both single- and multi-fidelity models show good convergence within a flux-driven transport solver. The big data validation approach allows efficient extensions of the training database, positioning GKNN-0 as a fast and accurate surrogate model of gyrokinetic turbulence.