Fast modeling of turbulent transport in fusion plasmas using neural networks

Abstract

DOI to the publisher's website.• The final author version and the galley proof are versions of the publication after peer review.• The final published version features the final layout of the paper including the volume, issue and page numbers. Link to publication General rightsCopyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.• Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain • You may freely distribute the URL identifying the publication in the public portal.If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the "Taverne" license above, please follow below link for the End User Agreement:

Progress and challenges in understanding core transport in tokamaks in support to ITER operations

“…Finally, fast neural network surrogate quasi-linear models are under development for scenario optimization and controloriented applications. This development leverages the enormous recent advances in machine learning techniques, and is based on a regression of quasi-linear transport models using training sets derived from targeted large-scale generation of model input-output mappings [92][93][94][95][96]. These faster versions will allow more large-scale validation and systematic scenario optimization, will facilitate core-pedestal coupled simulations towards full-device physics-based self-consistent scenario predictions, and will contribute to control-oriented applications.…”

Neural network surrogate of QuaLiKiz using JET experimental data to populate training space

“…This paper extends the previous NN approach [17] by incorporating the known impact of fuel dilution [18], plasma rotation [19] and magnetic equilibrium effects [20]. While several key aspects of these dependencies are already included in the previous work, referred to as the QLKNN-hyper-10D in this paper, their treatment is made more explicit by allowing the NN to learn from a more complete set of input variables within the context of the underlying model, QuaLiKiz.…”

“…This section provides a basic description of NNs and detail the NN architecture and training hyperparameters used for this study. In general, the NNs trained within this study used the same TensorFlow-1.6 training pipeline used by the previous work [17] and further details can be found in the referenced paper. The weights and biases of the NNs are publicly available in a GitLab repository, Ref.…”

“…QLKNN-hyper-10D used a lattice sampling method, referred to as a hyper-rectangular grid in Ref. [17], over 9 of its 10 input dimensions to populate its training dataset, with all of the input values carefully selected from domain expertise. The remaining parameter was applied via a reducedphysics model after NN evaluation.…”

Model order reduction approach to the one-dimensional collisionless closure problem

“…Real-time interpretation of experimental data, especially for future devices such as I is a strong motivation to create fast surrogate models based on quasilinear assumptions. Significant speedups have been obtained for instance by 43 using neural networks trained on a QuaLiKiz database. The same rational functions (57-59) can be used to accelerate the resolution of the gyrokinetic dispersion relation.…”