Speaker
Karel Lucas van de Plassche
Description
See the full Abstract at http://ocs.ciemat.es/EPS2018ABS/pdf/P2.1086.pdf
Real time capable turbulent transport modelling using neural networks
K. van de Plassche1 , J. Citrin1 , C. Bourdelle2 , V. Dagnelie1,3 , F. Felici4 , A. Ho1
1 DIFFER - Dutch Institute for Fundamental Energy Research, Eindhoven, the Netherlands
2 CEA, IRFM, F-13108, Saint-Paul-lez-Durance, France.
3 Utrecht University, Princetonplein 5, 3584 CC, Utrecht, The Netherlands.
4 EPFL-SPC, CH-1015, Lausanne, Switzerland.
Quasilinear gyrokinetic models are very successful in predicting particle and heat transport in
tokamaks, and in reproducing experimental profiles in wide parameter regimes. One such code
is QuaLiKiz, validated on JET, ASDEX-U and Tore-Supra discharges [1, 2, 3, 4]. While an
impressive six orders of magnitude faster than local nonlinear gyrokinetics, they are still too
slow for efficient scenario optimization and realtime applications.
A feed-forward neural network regression of QuaLiKiz was used in a successful proof-of-
concept [5, 6]. Such a network can be evaluated within a few microseconds, another six orders
of magnitude faster than the original model. These networks are tested and designed for the
RAPTOR rapid profile evolution code [7], but can also be used in other frameworks.
This current work is a major extension of the proof-of-principle from 4D to 10D. A large
database of 3.108 flux calculations over a 9D input space generated with the QuaLiKiz code is
used to extend the original input space of ion temperature gradient R/LTi , ion-electron temper-
ature ratio Ti /Te , safety factor q and magnetic shear ŝ with the electron temperature gradient
R/LTe , density gradient R/Ln , minor radius ρ, collisionality ν ∗ and effective ion charge Ze f f .
The 10th dimension, ExB shear, is added in post-processing using a new turbulence quenching
rule [8].
We present our methodology of training and validating these neural networks, which are
ready for applications within RAPTOR [9]. The speed of the networks created in this work
combined with RAPTOR allow for transport simulations at a speed that is unprecedented, and
opens new avenues in the modelling of fusion experiments.
References
[1] J. Citrin et al. Plasma Physics and Controlled Fusion 59 12400 (2017)
[2] C. Bourdelle et al., Plasma Physics and Controlled Fusion 58, 1 (2016)
[3] O. Linder et al., to be submitted to Nuclear Fusion
[4] M. Marin et al., this conference (EPS Prague 2018); A. Ho et al., this conference (EPS Prague 2018)
[5] J. Citrin et al. Nuclear Fusion 55 092001 (2015)
[6] F. Felici et al., submitted to Nuclear Fusion
[7] F. Felici and O. Sauter, Plasma Physics and Controlled Fusion 54, 2 (2012)
[8] V.I. Dagnelie et al. University of Utrecht MSc thesis (2017)
[9] F. Felici et al., this conference (EPS Prague 2018)
