Active learning pipeline for surrogate models of gyrokinetic turbulence

Model-based plasma scenario optimization often excludes reduced-order gyrokinetic models, e.g. QuaLiKiz, deeming them too slow for highly iterative applications. Previously, QLKNN, a feed-forward neural network (NN) surrogate model of QuaLiKiz, has shown a factor 10⁴ prediction speedup which enables its use in optimization. However, the training set generation still demanded considerable computational resources due to its size. Moreover, the QLKNN-jetexp-15D dataset only a minority of unstable points (<30% per turbulent mode) and may have oversampled regions. Such brute-force approaches are not feasible for more computationally-intensive models.

This study proposes an active learning (AL) method which uses output uncertainties to identify regions where additional data would improve the NN. A subset of the 15D dataset was used to train initial uncertainty-aware turbulent flux regression NNs and stable region classification NNs. New points for subsequent retraining iterations were selected via the NN uncertainties. A factor 250 data reduction was achieved while retaining an equivalent NN MSE. Further tests in integrated models are planned. This method shows promise in applicability to more accurate gyrokinetic simulations.