Refine
Year of publication
- 2024 (2) (remove)
Document Type
- Doctoral Thesis (2)
Language
- English (2)
Has Fulltext
- yes (2)
Is part of the Bibliography
- no (2)
Keywords
- Fusion (2) (remove)
Institute
Physics-regularized Machine Learning To Approximate 3D Ideal-MHD Equilibria At Wendelstein 7-X
(2024)
The magnetohydrodynamic (MHD) equilibrium model is one of the fundamental building blocks in the description of a magnetically confined plasma. The computational cost of constructing solutions to the 3D ideal-MHD equilibrium problem is one of the limiting factors in stellarator research and design; in particular, it limits the extent to which we can perform sample-intensive applications, applications which require many samples to be evaluated to yield meaningful results. Sample-intensive applications in stellarator research and design include, for example, equilibrium reconstruction, stellarator optimization, and flight simulators. In this thesis, we investigate how faithfully artificial neural networks (NNs) can quickly approximate ideal-MHD equilibria in stellarator geometries, starting with Wendelstein 7-X (W7-X), the world’s most advanced stellarator. In particular, we investigate (see section 1.7):
RQI: to what extent can NN models approximate the MHD equilibrium solution for different W7-X configurations and plasma profiles? What
is the speed-accuracy trade-off offered by NN models?
RQII: to what degree the NN model faithfully reproduces equilibrium quantities of interest (e. g., MHD stability)? To what extent can NN models meet the requirements of downstream applications (e. g., Bayesian
inference, stellarator optimization) in terms of equilibrium quantities
accuracy?
RQIII: whether we can exploit the implicit representation of a MHD equilibrium, i. e., the equilibrium solution should satisfy the ideal-MHD force
balance equation, to improve the NN approximation’s accuracy;
RQIV: the reconstruction of the full posterior istribution of plasma parameters and equilibrium quantities with self-consistent MHD equilibria; moreover, how does the adoption of MHD equilibria approximated by NN models affect the inferred plasma parameters?
A deep NN model is developed to learn the ideal-MHD solution operator in W7-X operational subspace, yielding 3D equilibria up to six orders of magnitude faster than currently available MHD equilibrium codes. Physics domain knowledge is embeded into the NN model: equilibrium solution symmetries are satisfied by construction, and the MHD force balance regularizes the NN model to satisfy the ideal-MHD equations. The model accurately predicts the equilibrium solution and it faithfully reproduces global equilibrium quantities and proxy functions used in stellarator optimization. Finally, the developed fast NN equilibrium model has been applied in downstream applications to obtain W7-X configurations with improved fast-particle confinement and to infer plasma parameters with self-consistent MHD equilibria at W7-X.
In course of the recent results from Wendelstein 7-X, stellarators are on the brink for assessing their maturity as a fusion reactor. To this end, stellarator specific transport regimes need detailed exploration both with appropriate systematic experimental investigations and models. A way to enhance the efficiency of this process is seen in an systematic evaluation of existing experimental data. We propose appropriate tools developed in information theory for examining large datasets. Information entropy calculations, that have proven to assist the systematic assessment of datasets in many other scientific fields, are used for novelty detection.
Potentially, as a first use-case of this holistic process, this thesis attempts to link and to develop approaches to examine the stellarator specific core-electron-root-confinement (CERC) regime. The specific interest for CERC emerges from the behavior of the radial electric field. While ion-root conditions exhibit negative radial electric fields, CERC’s positive field in the very core of fusion grade plasmas adds an outward thermodynamic force to high-Z impurities and could add to potential actuators to control impurity influx as to be examined for full-metal wall operation in large stellarators. Recently, this feature received revived intent for reactor scale stellarators.
Also, in this work, parameter regions close to the transition from ion-root to CERC are
examined. At lower rotational transform (a characteristic feature of the magnetic field confining fusion grade plasmas), transitions were detected when the plasma current evolved. As in smaller stellarators, it is concluded that low-order rationals and magnetic islands are related to the transitions. This is widely supported by extensive MHD simulations which finally provide indications for the role of zonal flow oscillations. As one of the outcomes, gyrokinetic instabilities are seen interacting for the first time with the neoclassical mechanisms in experiments.
In order to cope with the vast number of highly sampled spatio-temporal plasma data, new
techniques for novelty detection are required. Fundamental prerequisites for the detailed
physics investigations were the feasibility study of entropy-based data analysis techniques, and their adaptation to detect previously unrevealed transition mechanisms. These tools were applied to multivariate bulk plasma emissivity data, which allowed the exploration of large parameter spaces and provided insights in the spatio-temporal dynamics of CERC transitions.
In this manner, this research highlights the feasibility of information flow measure analysis in fusion studies. Applications of different entropy-based complexity measures are explored and this work sheds light on the capabilities, added value and limitations of these techniques. This investigation presents the integration of information flow measures to gain deeper understanding of plasma transport phenomena, by providing an approach to fast systematic data mining suited for real-time analysis. This work paves the way for further development and implementation of information-theoretic methods for plasma data analysis.
In summary, this research highlights the gained insight on CERC transitions, while showcasing the feasibility, added values and limitations of information flow measure analysis for fusion studies, to induce theory based analysis revealing new insights in fundamental, stellarator-specific transport mechanisms.