Regular Articles

Full Text / References

Development of a Surrogate Turbulent Transport Model and Its Usefulness in Transport Simulations

Mitsuru HONDA and Emi NARITA

National Institutes for Quantum and Radiological Science and Technology, Naka 311-0193, Japan

(Received 2 November 2020 / Accepted 15 November 2020 / Published 10 February 2021)

Abstract

For accelerating a transport simulation with an advanced physics turbulent transport model like TGLF, we have been developing a surrogate model that mimics the behavior of the model based on a neural network model. With a steady-state transport solver GOTRESS used, the surrogate model has shown its ability to successfully predict temperature profiles almost equivalent to those by TGLF. The performance of the surrogate model is improved by optimizing hyperparameters and eliminating outliers from training data. Extrapolability of the optimized model is examined by changing the normalized temperature gradient. The objective is to better investigate the nature of the model in addition to measuring its utility in transport simulations. The versatile model, which has been trained with data of multiple cases, is developed applicable to many situations. It shows the same reproducibility as the model specific to each individual case, a fact which unveils great potential of the surrogate model in transport simulations.

Copyright (c) 2021 The Japan Society of Plasma Science and Nuclear Fusion Research

Keywords

global optimization, neural network model, surrogate model, hyperparameter optimization, turbulent transport model, transport simulation, tokamak

DOI: 10.1585/pfr.16.2403002

Full Text

PDF format (640Kbytes)

References

• [1] M. Honda, Comput. Phys. Commun. 231, 94 (2018).
• [2] M. Honda and E. Narita, Phys. Plasmas 26, 102307 (2019).
• [3] G.M. Staebler, J.E. Kinsey and R.E. Waltz, Phys. Plasmas 12, 102508 (2005).
• [4] G.M. Staebler, J.E. Kinsey and R.E. Waltz, Phys. Plasmas 14, 055909 (2007).
• [5] O. Meneghini, C.J. Luna, S.P. Smith and L.L. Lao, Phys. Plasmas 21, 060702 (2014).
• [6] J. Citrin et al., Nucl. Fusion 55, 092001 (2015).
• [7] O. Meneghini et al., Nucl. Fusion 57, 086034 (2017).
• [8] E. Narita, M. Honda, M. Nakata, M. Yoshida, H. Takenaga and N. Hayashi, Plasma Phys. Control. Fusion 60, 025027 (2018).
• [9] E. Narita, M. Honda, M. Nataka, M. Yoshida, N. Hayashi and H. Takenaga, Nucl. Fusion 59, 106018 (2019).
• [10] J. Bergstra, R. Bardenet, Y. Bengio and B. Kégl, in Proceedings of the Advances in Neural Information Processing Systems 24 (NIPS 2011), Granada, Spain (2011), Vol.24.
• [11] http://jaberg.github.io/hyperopt/
• [12] http://maxpumperla.com/hyperas/
• [13] M. Abadi et al., “Tensorflow: Large-scale machine learning on heterogeneous distributed systems”, preprint arXiv:1603.04467 (2016).
• [14] K.L. van de Plassche et al., Phys. Plasmas 27, 022310 (2020).
• [15] J. Kirkpatrick et al., preprint arXiv:1612.00796 (2017).
• [16] C. Atkinson et al., “Pseudo-Recursal: Solving the Catastrophic Forgetting Problem in Deep Neural Networks”, preprint arXiv:1802.03875 (2018).
• [17] C. Atkinson et al., “Pseudo-Rehearsal: Achieving Deep Reinforcement Learning without Catastrophic Forgetting”, preprint arXiv:1812.02464 (2019).
• [18] G. Hinton, O. Vinyals and J. Dean, “Distilling the Knowledge in a Neural Network”, preprint arXiv:1503.02531 (2015).