Speaker
Description
The fast and reliable reconstruction of plasma X-ray and neutron emissivity is crucial in nuclear fusion devices for real-time monitoring of essential parameters, such as electron temperature, impurity concentration and ion fuel ratio. The ongoing research, conducted jointly by IRFM CEA (Cadarache, France) and IFJ PAN (Krakow, Poland), aims to develop, validate and implement machine learning methods - in particular evolutionary algorithms (e.g. genetic algorithms, GA) and artificial neural networks (NNs) – for this purpose. This contribution highlights the following recent efforts and results:
(1) Combining a GA with a customized Monte-Carlo (MC) code to optimize the design and performance of diagnostic systems for tokamak plasmas. The current case study focuses on a thin-foil proton-recoil system for neutron spectroscopy, involving intensive computations in a GA-MC framework and its validation with GEANT4 [1]. It is demonstrated that the GA-MC approach is well-adapted to such multi-dimensional optimization problem and that it can quantify the trade-off between geometrical parameters (e.g. converter thickness, detector dimensions, etc.) to optimize both the detection efficiency and energy resolution of the system. The proposed methodology is generic and may be applied in future work to the design optimization of other neutron or X-ray diagnostic systems resilient to neutron flux in the ITER environment.
(2) Employing fully-connected NNs to automate reconstruction of tungsten impurity concentration and distribution in the WEST plasma core (a crucial information to control the radiated power), using a large experimental training database from multiple diagnostics. Parametrization of the W profile has been recently introduced in the analysis, and the NN approach is validated for several WEST discharges against existing synthetic diagnostic tools based on forward modelling, showing a good consistency and a reduction of the computing time by orders of magnitude.
(3) Applying convolutional NNs to solve the inverse problem for X-ray tomographic reconstruction of the tokamak plasma emissivity field, which is essential to monitor impurity asymmetries. Since convolutional layers are adapted to techniques involving image processing, such architecture has been implemented for the SXR WEST tomographic system and compared with existing methods involving Tikhonov regularization and fully-connected NN. Very promising results were obtained, in particular lower reconstruction error and higher spatial resolution than fully-connected NN with a computing time still several order of magnitude lower than Tikhonov regularization.
These results open the door to significantly enhanced capabilities for reconstructing plasma parameters, particularly to study ionization, transport and radiation properties of heavy impurities in tokamaks such as WEST or ITER with a substantial suprathermal electron population.
[1] S. Agostinelli et al., Nucl. Instrum. Meth. A 506 (2003) 250-303.