Detection and prediction of a beam-driven mode in field-reversed configuration plasma with recurrent neural networks
P. 1

International Atomic Energy Agency Nuclear Fusion Nucl. Fusion 60 (2020) 126025 (11pp) https://doi.org/10.1088/1741-4326/abb328
Detection and prediction of a beam-driven mode in field-reversed configuration plasma with recurrent neural networks
Cory Scott1, Sean Dettrick2, Toshiki Tajima2,3, Richard Magee2 and Eric Mjolsness1
1 Dept. of Computer Science, University of California, Irvine, CA, United States of America 2 TAE Technologies, Inc, Foothill Ranch, CA, United States of America
3 Dept. of Physics, University of California, Irvine, CA, United States of America
E-mail: scottcb@uci.edu
Received 4 May 2020, revised 11 August 2020 Accepted for publication 27 August 2020 Published 14 October 2020
Abstract
Energetic beams excite semi-repetitive modes (‘staircase mode’) in the field-reversed configuration (FRC) plasma. We explore several neural network architectures to detect, and in some cases predict, this type of mode onset. We weigh the performance of these architectures and find that recurrent neural networks (RNNs), specifically long short-term memory (LSTM) networks, outperform all other models we examine. LSTMs can predict the onset of staircase with a lead window of 0.2 ms, which has implications for plasma longevity and is a promising direction for similar analysis in FRC devices in the future.
Keywords: field-reversed configuration, fusion, plasma physics, machine learning, convolutional neural networks, recurrent neural networks
 (Some figures may appear in colour only in the online journal)
1. Background
In the magnetized plasma of a fusion reactor we often observe instabilities and disturbances. These may be avoidable by adjusting the initial or boundary conditions of the plasma to keep it in a maintainable and controllable state. Because these instabilities may be the result of dynamic plasma behavior, it is desirable to develop a set of methods or strategies to either predict and / or adjust the evolving plasma conditions in real- time. The rapid evolution of machine learning (ML) methods in other types of plasma fusion settings, and indeed plasma fusion control [1], suggests that it may be fruitful to paramet- rize this kind of controller using an artificial neural network (ANN). In that example, predictive time for an ANN was on the order of several ms, which could be sufficient time (in the future) for the model ANN to respond to and adjust the plasma in a tokamak, in real time. Following such examples, here we attempt to predict disturbances within a field-reversed config- uration (FRC) plasma, in a device called C-2U.
In the C-2U device, two compact tori are formed by the theta pinch method, and accelerated towards each other to col- lide and merge, forming a field reversed configuration (FRC) [2]. An energetic ion population created by neutral beam injec- tion (NBI) into the FRC is found experimentally to provide sufficient heating, current drive, and kinetic stabilization to increase the FRC lifetime from the 1 ms achieved without NB to 10 ms or higher [3]. This timescale exceeds that of any known macroscopic (MHD type) instabilities which could potentially degrade the FRC confinement. Thus, the FRC may stay in a regime free of macro-instabilities. Nevertheless, with some combinations of NB and FRC parameters, kinetic instabilities are observed.
The micro-burst instability [4] is characterized by small, semi-periodic drops in internal pressure due to a redistribu- tion of fast ions which results in a staircase-like pattern in the time trace of the midplane excluded flux radius and gives the mode its nickname, “the staircase” (see figure 1). This drop in internal pressure results in a steepening of the plasma density
 1741-4326/20/126025+11$33.00 1
© 2020 IAEA, Vienna Printed in the UK


















































































   1   2   3   4   5