Active learning of turbulent transport predictions towards robust generalized surrogate models

High-fidelity plasma turbulent transport models are often not included in integrated plasma modelling (IM) predictions due to their computational cost. Neural network (NN) regressions of reduced models, e.g. QuaLiKiz [1] and TGLF [3], can speed up the prediction while keeping comparable accuracy [4]. Extending this to higher-fidelity models requires minimizing training set requirements and limiting biases with machine-specific sampling. Recent work with an active learning (AL) approach on a QuaLiKiz dataset yielded a factor 10 less data to generate a surrogate model of equivalent performance, named the ADEPT (Active Deep Ensembles for Plasma Turbulence) model [5]. This study extends the AL study by using interpretable uncertainty-aware NN architectures [6] and improved batch acquisition functions [7]. The initial study was performed with QuaLiKiz to assess improvements, though extensions to TGLF and potentially CGYRO are foreseen. In addition, the built-in uncertainty estimation can be used within IM for identifying untrained parameter spaces and/or for uncertainty quantification exercises.

[1] K. L. van de Plassche, Physics of Plasmas 27, 022310 (2020)

[2] O. Meneghini et al., Nuclear Fusion 57, 086034 (2017)

[3] A. Ho et al., Nuclear Fusion 63, 066014 (2023)

[4] L. Zanisi et al., Nuclear Fusion 64, 036022 (2024)

[5] A. Panera Alvarez et al., arXiv:2402.00760 (2024)

[6] A. Kirsch et al., arXiv:2302.08981 (2023)