An active learning pipeline for surrogate models of gyrokinetic turbulence

Digital twinning of a tokamak device requires fast system state inference. Physics-based computational models that predict future states are often too slow to be actionable, and thus undesirable for offline scenario planning. These tasks may be performed faster if the physics-based model is replaced by a neural network-based surrogate. Obtaining the labels to train the surrogate can be computationally expensive, additionally, some inputs may result in trivial outputs. Here we propose a two-stage active learning pipeline for digital twinning of gyrokinetic turbulence in the core of tokamak fusion plasmas. Our pipeline leverages an uncertainty-based acquisition function which greatly outperforms random acquisition and leads to a reduction of 99.6% in the amount of input-output mappings needed from the physical model without compromising on performance.