A machine learning framework for determining quasilinear saturation rules of turbulent transport from linear gyrokinetic data

Turbulent instabilities are a primary driver of transport in the core of tokamak plasma. Nonlinear gyrokinetics represents the most accurate plasma turbulence modelling paradigm available to accurately simulate local turbulent fluxes. However, full fidelity gyrokinetic simulations are computationally expensive making it impractical for use in integrated modelling simulations. Quasilinear models are used for this purpose where the linear response of the plasma instabilities is combined with an estimation of the magnitude of the saturated potentials via a saturation rule to provide calculations in reduced time.



The latest quasilinear saturation rule, SAT3, is an accurate model of the saturated potential for an enhanced CGYRO database that encompasses most common ITG and TEM turbulence regimes found in experiments. It does not however represent a complete description of turbulent plasma transport and will perform less reliably in a parameter space far from its empirical regression. In this work, we formulate a generalized framework to predict the magnitude of the 1D saturated potentials against the poloidal wavenumber from local linear simulation, growth rates and frequencies, using a neural network architecture trained on a database of high-resolution nonlinear CGYRO simulations. A machine learning fit that models the squared potentials more accurately will improve the calculation of the total energy and particle fluxes as well as considerably reduce the simulation time in integrated modeling.