Towards the Development of a Neural-Network Version of NUBEAM for Scenario Optimization and Control in KSTAR
S. Morosohk, M.D. Boyer, E. Schuster
KSTAR Conference
Seoul, South Korea, February 20-22, 2019
NUBEAM is a high-accuracy simulation code developed to calculate the
effects of neutral beams on the plasma. However, the code is
computationally demanding due to its Monte Carlo nature. This makes it
unrealistic to use in real-time control or in situations where a large
number of simulation runs are required, such as in model-based scenario
optimization for instance. A neural network model (NubeamNet) has been
originally developed to closely approximate NUBEAM results for NSTX-U
at drastically reduced run time [1]. More recently, a database of NUBEAM
runs has been developed for DIII-D and used to train a neural network
for model-based applications requiring a faster run time. The trained
neural network can match the NUBEAM data with high accuracy, and requires
only a fraction of a second to compute the outputs for a given set of
inputs. The trained model is suitable for integration into real-time
control systems and for use in model-based scenario planning. Simulation
results are presented to illustrate the potential of the trained
neural-network version of NUBEAM in DIII-D. Since the same neural network
can be retrained for different machines, present and planned work toward
extending this technique to KSTAR will be discussed. It is anticipated
that this machine-learning-based model for the neutral-beam-generated
heating, torque, and current drive may play an important role in the
development of algorithms for scenario optimization and control in KSTAR.
[1] M.D. Boyer et al., “Real-time capable neural network approximation
of NUBEAM for use in the NSTX-U control system,” 45th EPS Conference
on Plasma Physics, 2018.