Speaker
Axel Jardin
Description
See the full Abstract at http://ocs.ciemat.es/EPS2018ABS/pdf/P4.1013.pdf
Neural networks for fast soft X-ray tomographic inversions in tokamaks
A. Jardin1, J. Bielecki1, D. Mazon2, J. Dankowski1, Y. Peysson2 and M. Scholz1
1
Institute of Nuclear Physics Polish Academy of Sciences (IFJ PAN), PL-31-342, Krakow,
Poland
2
CEA, IRFM F-13108 Saint Paul-lez-Durance, France
Reconstructing the local plasma emissivity in the soft X-ray (SXR) range 0.1 – 20 keV can be
very useful to access essential information on particle transport, magnetohydrodynamic activity
or impurity content in tokamaks. In particular, radiative cooling of heavy impurities like
tungsten (W) could be detrimental for the plasma core performances of ITER and developing
robust SXR diagnostics is thus a crucial issue for monitoring the impurities and the prevention
or their central accumulation. 2D tomography is the usual method to access the local SXR
emissivity from line integrated measurements of two or more cameras viewing the plasma in a
poloidal cross-section. This mathematically ill-posed and quite challenging problem is
traditionally solved using Tikhonov regularization, such as the minimum Fisher information
method implemented on the ASDEX Upgrade, TCV and WEST tokamaks [1]. However,
real-time impurity control will require fast inversion methods while Tikhonov regularization
needs relatively significant computational time, although many efforts have been performed in
the direction of real-time applications [2].
Neural networks have been recently used for plasma tomography with JET neutron camera [3]
and bolometers [4], showing promising results in terms of computational time as well as quality
of reconstruction. Thus, the aim of this contribution is to investigate the use of neural network
for fast soft X-ray tomographic reconstructions in the prospect of real-time impurity control in
tokamak plasmas. Using a large database of various synthetic emissivity phantoms to train the
neural network, the robustness of the inversion will be investigated and benchmarked with
traditional Tikhonov regularization techniques [5]. The influence of the network parameters
(neuron activation function, hidden layers, regularization procedure…) will be studied with a
specific attention to the net gain in computational time, to the robustness of the method with
respect to the experimental noise in the measurements and to the intrinsic limitations of such
approach.
Acknowledgments. This work has received a financial support from the POLONIUM collaboration program
between the French and Polish Ministers of Science.
[1] A. Jardin et al, 2016 JINST 11 C07006.
[2] J. Mlynar, V. Weinzettl, G. Bonheure, A. Murari, Fusion Sci. Technol. 58 (3) (2010) 733–741.
[3] E. Ronchi et al, Nuclear Instruments and Methods in Physics Research A 613 (2010) 295–303.
[4] F. Matos, D. Ferreira, P. Carvalho and JET contributors, Fusion Engineering and Design 114 (2017) 18–25.
[5] A. Jardin, D. Mazon and J. Bielecki, Phys. Scr. 91 (2016) 044007.
