Control-Oriented Current-Profile Response Modeling Using Neural Network Accelerated Versions of TGLF and NUBEAM for DIII-D
S. Morosohk, T. Rafiq, E. Schuster, O. Meneghini, D. Boyer
Division of Plasma Physics (DPP) Annual Meeting of the American Physical Society (APS)
Remote, November 9-13, 2020
|
Abstract
|
|
Model-based control methods for robust current-profile regulation rely
on control-oriented as opposed to physics-oriented models. These
typically use a number of limiting assumptions to be able to achieve
the necessary calculation speeds. The evolution of the current profile
depends both on the plasma resistivity, which is primarily a function
of the temperature, and on the deposition characteristics of the heating
and current drive sources. Recent work [1, 2] has shown success with
the use of neural networks to recreate the neutral-beam heat and current
depositions computed by the Monte Carlo module NUBEAM and the heat and
particle fluxes computed by the quasilinear transport model TGLF. These
neural network models can be run in CPU-microseconds, enabling the
possibility of real-time prediction for profile control purposes. In
this work, the electron heat flux profile computed by the TGLF neural
network and the neutral-beam heat and cur- rent depositions computed
by the NUBEAM neural network are integrated into the magnetic diffusion
and electron heat transport equations within COTSIM (Control- Oriented
Transport Simulator). This eliminates the need for empirical correlations
for the electron temperature and the neutral-beam heat and current
depositions.
[1] S. Morosohk, M.D. Boyer and E. Schuster, APS-DPP 2018.
[2] O. Meneghini et al. 2017 Nuclear Fusion 57 086034.
