Two-level Bayesian Optimization for Steady-State Predictions in Gyrokinetic Simulations of Plasma Turbulence

Flux matching for predicting steady states is a key challenge in turbulent gyrokinetic simulations, which model plasma behavior in fusion devices such as tokamaks. This task can be posed as an optimization problem: identifying the background plasma gradients that minimize the mismatch between turbulent fluxes from gyrokinetic simulations and those predicted by a transport model that includes heating and loss mechanisms. We propose a multifidelity Bayesian optimization approach for solving the flux matching problem.

This approach has been implemented using the PORTALS component of MITIM, an open-source resource developed for plasma modeling applications. The talk will demonstrate its application using information from two versions of the TGLF reduced turbulence model with differing saturation rules. We explore how combining these sources of information can lower the computational cost of the optimization process.