Data-driven study of major disruption prediction and plasma instabilities across multiple tokamaks
| Autor: | Zhu, Jinxiang |
|---|---|
| Rok vydání: | 2023 |
| Druh dokumentu: | Diplomová práce |
| Popis: | The use of nuclear fusion energy via magnetic-confinement tokamaks is one of a few encouraging paths toward future sustainable energy. Along the way, scientists need to learn to avoid plasma disruptions: these sudden and unexpected plasma terminations still represent one of the key challenges for tokamak devices. Forecasting plasma instabilities and disruptions using first-principle models has been demonstrated to be extremely difficult, due to the complexity of the problem and the high non-linearity of the system. To date, disruption and plasma instabilities prediction has been studied through two main approaches: data-driven versus physics-driven (or model-based). On the one hand, recent statistical and machine learning (ML) approaches based on experimental data have shown attractive results for disruption prediction, even in real-time environments. Different tokamak devices have different operational spaces, spatiotemporal scales for physics events, and plasma diagnostics. Therefore, most of these data-driven approaches were developed and optimized specifically for one device and did not show promising cross-device predictive ability. In addition, the complexity of these data-driven models limits their physics interpretability. Recent Deep-Learning (DL) based disruption prediction studies demonstrate the potential for acquiring a general representation of experimental data that can be used in cross-machine applications. On the other hand, model-based studies seek to identify event chains that can lead to disruptions through early event detection, which can help operators to avoid plasma instabilities disruptions. However, the extrapolation ability of physics-based models to new devices, especially to new physics regimes is still unclear. This thesis demonstrates the application of data-driven methods on plasma insta-bilities and disruption prediction via four major contributions. First, through explo-rative data analysis of thousands of shots on C-Mod, DIII-D and EAST tokamaks, the advantage of sequence-based disruption prediction model was shown. Based on this finding, a new Hybrid Deep-Learning (HDL) general disruption predictor was developed using C-Mod, DIII-D and EAST databases and it achieves state-of-the-art performance on three machines with only limited hyperparameter tuning. Dedicated cross-machine disruption prediction studies using this HDL model demonstrated that a significantly boosted accuracy on the target machine was achieved by training on 20 disruptive shots, thousands of non-disruptive shots from the target machine com-bined with hundreds of disruptive shots from other devices. In addition, by comparing the predictive performance of each individual numerical experiment, the disruptive shots from multiple devices were found to contain device-independent knowledge that can be used to inform predictions for disruptions occurring in a new device while non-disruptive shots were found to be machine-specific. Second, the cross-regime disruption prediction on multiple tokamaks using HDL model demonstrated data-driven disruption predictors trained on abundant Low Performance (LP) discharges work poorly on the High Performance (HP) regime of the same tokamak, which is a consequence of the distinct distributions of the tightly correlated signals related to disruptions in these two regimes. Moreover, the cross machine experiments suggested matching operational parameters among tokamaks strongly improves cross-machine accuracy. Given these conclusions, a scenario adaptive strategy that works for all data-driven models was proposed for next generation tokamaks, such as ITER and SPARC, and highlight the importance of developing baseline scenario discharges of future tokamaks on existing machines to collect more relevant disruptive data. Third, the powerful HDL model was upgraded to an integrated ML model that can predict major disruption as well as multiple unstable events in tokamak plasmas that can facilitate the physics interpretation of output from the black box data-driven models and enables disruption avoidance by responding to early unstable events of plasmas. Enhanced cross-machine ability and improved warning time was also observed using the integrated ML model. Finally, among all different plasma unstable events, the 𝑛 = 1 tearing mode (TM) is considered to be one of the most important disruption precursors and its predictive ability is strongly desirable for ITER and SPARC. In the final part of this thesis, an empirical boundary for the 𝑛 = 1 tearing mode (TM) is developed via data-driven methods and verified on thousands of DIII-D discharges. The fitted boundary is a linear function of plasma equilibrium parameters such as collisionality, poloidal beta, and the MHD risk factor (a combination of the normal-ized electron temperature profile width, q95 and elongation). The boundary indicates with a value related to the probability of having the TM onset and it achieves 88% of shot-by-shot accuracy in offline analysis of DIII-D data. Preliminary cross-machine analysis of TM onset prediction shows potential applicability of the empirical bound-ary to C-Mod and EAST data as well, but the relative importance of the individual parameters is different for different devices. This suggests the existence of different trigger mechanisms for the TMs, implying that the boundary could be generalized using data from different tokamaks representing different trigger mechanisms to im-prove its extrapolability. Finally, this new proximity metric to the 𝑛 = 1 TM onset has been incorporated into the real-time in DIII-D plasma control system (PCS) and results from real-time experiments will be discussed. Ph.D. |
| Databáze: | Networked Digital Library of Theses & Dissertations |
| Externí odkaz: |
|