Building confidence in data-driven surrogate transport models for turbulent plasmas
Résumé
Getting fast and reliable predictions of turbulent transport properties is an important challenge in magnetic fusion. Previous research [R. A. Heinonen & P. H. Diamond, Phys. Rev. E 101 061201 (2020)] proposed a data-driven approach using neural networks to predict the particle flux and Reynolds stress in a minimal model of drift-wave turbulence. The present work extends this approach to the interchange instability driven by the magnetic curvature. An assessment of the limits and caveats associated with a data-driven approach based on machine learning regression algorithms is performed -an essential step for scalability toward more complex high-fidelity codes. In particular, a figure of merit is introduced to indicate regions within the parameter space where the neural network outputs can be trusted. Some applications of the data-driven surrogate model are presented. Specifically, predictions are used to gain insight into the vorticity gradient's contribution to the turbulent flux and the antiviscous nature of the Reynolds stress.
Origine | Fichiers produits par l'(les) auteur(s) |
---|