Surrogate model generation using high-fidelity CGYRO predictions enabled by active learning techniques

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, 2] and TGLF [3], can speed up the prediction while keeping comparable accuracy [4]. However, the standard architectures and training techniques are known to require large training datasets, whereas the data availability for high-fidelity gyrokinetic models, such as CGYRO [5], are limited by computational bandwidth. Recent work with an active learning (AL) approach on QuaLiKiz yielded a performance comparable to the original network with a factor 10 less data [6] and decent performance under even tighter dataset size restrictions [7]. This study investigates the application of the developed AL pipeline to CGYRO within the parameter space of SPARC, which efficiently automates the input sampling process based on information gain. Removal of the turbulent mode separation results in poorer predictions at mode transition boundaries but may not be catastrophic to the SPARC parameter space due to the dominance of ITG turbulence.

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

[2] A. Ho et al., Physics of Plasmas 28, 032305 (2021)

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

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

[5] J. Candy et al., Journal of Computational Physics 324, 73-93 (2016)

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

[7] A. Ho et al., Physics of Plasmas, (submitted 2025)