Page 10 - Detection and prediction of a beam-driven mode in field-reversed configuration plasma with recurrent neural networks

Nucl. Fusion 60 (2020) 126025
C. Scott et al
6. Conclusion and future work
We examine several ANN models for analyzing the staircase mode observed [4] in NB-driven C-2U FRCs. Using an optim- ized LSTM, we are able to classify this instability to a moder- ate degree of accuracy ( 85%). The LSTM structure is also able to perform regression on the excluded flux radius, R∆φ, the signal which most clearly shows staircase instability in these experiments, with a prediction lead time of 0.2 ms. We hope to improve accuracy by expanding this experiment to a larger dataset, which will likely decrease the amount of overfitting. In addition, we hope to repeat the variable-importance exper- iment, with more trials, on this larger dataset.
Since performing this study the successor to the C-2U device, C-2 W, has commenced operation and had a success- ful experimental campaign [29]. This machine has a signific- antly larger and higher fidelity diagnostic suite [30] and, with its machine-readable Machine State Database (MSDB) it is an ideal target of AI research. We intend to apply the methods developed here to that new device. In particular, there are many more shots (at least 4000) available for study from C-2 W. More shots, as well as a wider variety of signals, will facilit- ate repeating the variable importance study on a much larger scale, up to and including an exhaustive search over input sig- nals, rather than random trials.
Finally, a major goal is increasing the lead time of the pre- diction detailed in section 5. We have demonstrated that the network can predict an instability reasonable accurately up to 0.2 ms before it occurs. Increasing the amount of warning time is vital for avoiding this type of instability via a feedback mechanism. Our prediction time of 0.2 ms is at least an order of magnitude longer than the frequency of many of the phys- ical processes which are suspected to play a role in plasma instability (e.g. the plasma rotation frequency ( 0.02 ms), the microburst fluctuation frequency ( 0.01 ms), and the fast ion orbit frequency ( 0.001 ms)). In particular, recall from section 1 that neutral beams are one of the main control mechanisms of the C-2U device. The time required to turn NBI on or off, or to change their energy by 10-20 kV, is much less than 100 μs (see [31], figure 5), making our lead time of 0.2 ms (twice as much) a plausible element of a control loop to manage flux radius. Our work is therefore a useful mechanism for real-time monitoring of plasma stability and a vital step towards increas- ing plasma longevity.
Acknowledgment
This work was performed as part of the Summer Internship Program at TAE Technologies, Inc. We would like to thank the TAE team for their support and contributions. We are also grateful to the TAE shareholders for their continued support. This work was also supported in part by NSF NRT Award Number 1633631.
ORCID iDs
Cory Scott  https://orcid.org/0000-0002-5561-2368
Eric Mjolsness  https://orcid.org/0000-0002-9085-9171 References
[1] Hernandez J.V., Vannucci A., Tajima T., Lin Z., Horton W. and McCool S. 1996 Neural network prediction of some classes of tokamak disruptions Nucl. fusion 36 1009
[2] Guo H.Y. et al 2011 Formation of a long-lived hot field reversed configuration by dynamically merging two colliding high-β compact toroids Phys. Plasmas
18 056110
[3] Binderbauer M.W. et al 2015 A high performance field-reversed configuration Phys. Plasmas 22 056110
[4] Deng B.H. et al 2018 First experimental measurements of a new fast ion driven micro-burst instability in a field-reversed configuration plasma Nucl. Fusion
58 126026
[5] Thompson M.C. et al 2016 Diagnostic suite of the C-2U advanced beam-driven field-reversed configuration plasma experiment Rev. Sci. Instrum. 87 11D435
[6] Cannas B., Fanni A., Marongiu E. and Sonato P. 2003 Disruption forecasting at JET using neural networks Nucl. Fusion 44 68–76
[7] Pautasso G. et al 2002 On-line prediction and mitigation of disruptions in ASDEX Upgrade Nucl. Fusion 42 100–8
[8] Matos F. et al 2020 Classification of tokamak plasma confinement states with convolutional recurrent neural networks Nucl. Fusion 60 036022
[9] Canny J. 1986 A computational approach to edge detection
IEEE Trans. Pattern Anal. Mach. Intell. 8 679–98
[10] Jolliffe I.T. 1986 Principal components in regression analysis
Principal Component Analysis (Berlin: Springer) pp 129–55 [11] Pedregosa F. et al 2011 Scikit-learn: Machine learning in
python J. Mach. Learn. Res. 12 2825–30
[12] Abadi M. 2016 Tensorflow: large-scale machine learning on
heterogeneous distributed systems arXiv:1603.04467 [13] Kingma D.P. and Ba J. 2014 Adam: A method for stochastic
optimization arXiv:1412.6980
[14] Hornik K., Stinchcombe M. and White H. 1989 Multilayer
feedforward networks are universal approximators Neural
Netw. 2 359–66
[15] Haussler D. 1988 Quantifying inductive bias: AI learning
algorithms and valiant’s learning framework Artif. Intell.
36 177–221
[16] Bishop C.M. 2006 Pattern Recognition and Machine Learning
(Berlin: Springer)
[17] Goodfellow I., Bengio Y. and Courville A. 2016 Deep
Learning (Cambridge, MA: MIT Press)
[18] Rumelhart D.E., Hinton G.E. and Williams R.J. 1986 Learning
representations by back-propagating errors Nature
323 533–6
[19] Shalev-Shwartz S. and Ben-David S. 2014 Understanding
Machine Learning: From Theory to Algorithms
(Cambridge: Cambridge university press)
[20] Srivastava N., Hinton G., Krizhevsky A., Sutskever I. and
Salakhutdinov R. 2014 Dropout: A simple way to prevent neural networks from overfitting J. Mach. Learn. Res.
15 1929–58
[21] Hochreiter S. and Schmidhuber J. 1997 Long short-term memory Neural Comput. 9 1735–80
[22] Piczak K.J. 2015 Environmental sound classification with convolutional neural networks 2015 IEEE 25th Int. Workshop on Machine Learning for Signal Processing (MLSP) IEEE pp 1–6
[23] Vannucci A., Oliveira K.A. and Tajima T. 1999 Forecast of TEXT plasma disruptions using soft X rays as input signal in a neural network Nucl. Fusion 39 255
10