Speaker
Federico Felici
Description
See the full Abstract at http://ocs.ciemat.es/EPS2018ABS/pdf/P4.1091.pdf
Real-time multichannel tokamak plasma profile simulations using the
RAPTOR code and the QLK-NN first-principle transport model
F. Felici1 , J. Citrin2 , K. van de Plassche2 , A.A. Teplukhina1 , A. Ho2 , C. Bourdelle3 ,
O. Sauter1 , the EUROFusion MST1 Team4 and JET contributors5
1 EPFL-SPC, CH-1015, Lausanne, Switzerland. 2 DIFFER, Eindhoven, The Netherlands.
3 CEA, IRFM, France. 4 See http://www.euro-fusionscipub.org/mst1
5 See the author list of: X. Litaudon et al 2017 Nucl. Fusion 57 102001
Real-time capable yet accurate simulations of the plasma profile evolution have important ap-
plications in discharge preparation and optimization, real-time profile reconstruction and con-
trol.For the first time, real-time-capable coupled simulations of the kinetic profiles Te , Ti and ne
and q profile have been obtained that agree well with experimental data. The RAPTOR rapid
profile evolution code was used for this purpose, coupled to a first-principle based model to pre-
dict the transport coefficients for the kinetic profiles. This transport model, named QLK-NN, is
a neural network emulation of results from the QuaLiKiz quasilinear gyrokinetic code [1]. This
allows the fluxes to be evaluated in less than a millisecond per radial point per time step. An
initial version of QLK-NN, named QLK-NN4Dkin, extending the original proof of principle in
[2] to include kinetic electrons, takes as inputs the normalised logarithmic ion temperature gra-
dient R/LTi , the ion to electron temperature ratio Ti /Te , the safety factor q, and magnetic shear
s and returns the electron and ion heat flux, and electron diffusion and pinch coefficients. This
model was trained in the regime where turbulence drive is dominated by ITG modes and has
proven successful in reproducing kinetic profiles from a well-diagnosed JET discharge by set-
ting only the kinetic profile boundary conditions at the top of the pedestal [3]. First results using
RAPTOR with a more advanced version of the neural network transport model, QLK-NN10D
[4], are also presented. This model is trained on an extended parameter space including further
dependence on the electron temperature gradient, the density gradient, local aspect ratio, colli-
sionality, impurity content, and perpendicular flow shear, and covers ITG/TEM/ETG turbulence
regimes. First experimental tests are foreseen on JET and MST1 devices in 2018.
References
[1] Citrin, J. et al. 2017 Plasma Physics and Controlled Fusion 59 124005
[2] Citrin, J. et al. 2015 Nuclear Fusion 55 092001
[3] Felici, F. et al. 2018 Submitted to Nuclear Fusion
[4] v.d. Plassche, K. et al. This conference
