Multifidelity Bayesian Optimization for Steady-State Predictions using Gyrokinetic Simulations of Plasma Turbulence
Please login to view abstract download link
The flux matching problem is an important consideration in turbulent gyrokinetic simulations, which are used to study the dynamics of a plasma in fusion devices like tokamaks. By reformulating the flux matching problem (required to predict steady state in plasma systems) as an optimization problem, the problem reduces to finding the optimal background plasma gradients that minimize the discrepancy between the turbulent fluxes computed by the gyrokinetic simulation and the transport model, which considers volumetric heating and losses. We present a multifidelity Bayesian optimization approach for solving the flux matching problem. Our multifidelity approach uses information from 3 different sources: an analytic model of transport (low-fidelity), a reduced model of turbulence (medium-fidelity) and a full gyrokinetic turbulence code (high-fidelity). We investigate how information from these different sources can be combined to reduce the overall computational cost of the optimization procedure.