東華大學圖書館 |

Language: English

Help

回圖書館首頁

手機版館藏查詢

Back

Switch To: Labeled | MARC Mode | ISBD

Enhanced Plasma Profile Estimation a...

Morosohk, Shira.

Linked to FindBook

Google Book

Amazon

博客來

Enhanced Plasma Profile Estimation and Control in Tokamaks via Machine Learning.

Record Type:	Electronic resources : Monograph/item
Title/Author:	Enhanced Plasma Profile Estimation and Control in Tokamaks via Machine Learning./
Author:	Morosohk, Shira.
Published:	Ann Arbor : ProQuest Dissertations & Theses, : 2024,
Description:	297 p.
Notes:	Source: Dissertations Abstracts International, Volume: 85-07, Section: B.
Contained By:	Dissertations Abstracts International85-07B.
Subject:	Mechanical engineering. -
Online resource:	https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30687905
ISBN:	9798381377293

Enhanced Plasma Profile Estimation and Control in Tokamaks via Machine Learning.
Morosohk, Shira.

Enhanced Plasma Profile Estimation and Control in Tokamaks via Machine Learning. - Ann Arbor : ProQuest Dissertations & Theses, 2024 - 297 p.

Source: Dissertations Abstracts International, Volume: 85-07, Section: B.

Thesis (Ph.D.)--Lehigh University, 2024.

This item must not be sold to any third party vendors.

The tokamak concept is currently one of the most promising avenues to achieve energy generation through nuclear fusion, a feat that would enable the world to produce nearly limitless clean electricity. Unfortunately, no tokamak to date has ever achieved the conditions that would allow more energy to be generated from fusion reactions than it takes to sustain a plasma in that state. In order to produce and maintain a plasma that is both favorable for fusion reactions to occur and stable, a large number of different plasma properties must be carefully controlled. These plasma properties are all closely interconnected and display highly nonlinear behavior. In addition, a limited number of actuators are available that each produce multiple different effects, making it virtually impossible to design separate controllers for each plasma property that can operate independently. Instead, control solutions need to be developed that consider the interconnectedness of the system and use the same actuators to regulate multiple different plasma properties simultaneously. In order to quantify the interconnectedness of the system, model-based control techniques can be used that rely on predictive models to describe the plasma evolution.A number of different factors determine if a predictive model is suitable for control applications. The most important requirement of these models is usually the ability to run fast enough for the relevant application; the calculation speed requirement is often on the order of milliseconds. In order to achieve these calculation speeds, many physics-based control-oriented models make simplifying assumptions, sacrificing some of their accuracy. Empirical models can achieve very high levels of accuracy at fast enough calculation speeds, but can be limited in the range of plasma scenarios they are valid for. Machine learning offers a solution to these trade-offs: by training a machine learning algorithm to replicate the calculations of a slow, high fidelity physics-oriented code, a model can be developed that runs fast enough to be useful for control applications while retaining most of the accuracy of the high fidelity code and validity across a wide range of plasma scenarios. In this dissertation, two neural network surrogate models are trained to replicate the results of physics-oriented codes: NubeamNet predicts the effects of neutral beam injection on the plasma, and MMMnet predicts the turbulent diffusivity coefficients. These neural network surrogates are integrated with conventional models to improve the fidelity of the control-oriented predictive simulation code COT-SIM. This combination of machine learning-based and conventional models are then applied to a number of different model-based control applications. A feedforward optimization scheme that uses COTSIM including neural networks as its predictive model is developed to aid in scenario planning activities. An observer algorithm is devised to estimate the state of the electron temperature profile in real time, and has been validated in real time on the DIII-D tokamak. A feedback controller is designed to actively regulate the electron temperature profile, and is shown to successfully track a temperature profile target in experiment. Another controller is developed to actively track both the electron temperature profile and the safety factor profile simultaneously.

ISBN: 9798381377293Subjects--Topical Terms:

649730
Mechanical engineering.
Subjects--Index Terms:

Control systems

Enhanced Plasma Profile Estimation and Control in Tokamaks via Machine Learning.
LDR:04578nmm a2200385 4500 001 2395774
005 20240517105014.5
006 m o d
007 cr#unu||||||||
008 251215s2024 ||||||||||||||||| ||eng d
020 $a 9798381377293
035 $a (MiAaPQ)AAI30687905
035 $a AAI30687905
040 $a MiAaPQ $c MiAaPQ
100 1 $a Morosohk, Shira. $0 (orcid)0000-0002-3133-5095 $3 3765281
245 1 0 $a Enhanced Plasma Profile Estimation and Control in Tokamaks via Machine Learning.
260 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2024
300 $a 297 p.
500 $a Source: Dissertations Abstracts International, Volume: 85-07, Section: B.
500 $a Advisor: Schuster, Eugenio.
502 $a Thesis (Ph.D.)--Lehigh University, 2024.
506 $a This item must not be sold to any third party vendors.
520 $a The tokamak concept is currently one of the most promising avenues to achieve energy generation through nuclear fusion, a feat that would enable the world to produce nearly limitless clean electricity. Unfortunately, no tokamak to date has ever achieved the conditions that would allow more energy to be generated from fusion reactions than it takes to sustain a plasma in that state. In order to produce and maintain a plasma that is both favorable for fusion reactions to occur and stable, a large number of different plasma properties must be carefully controlled. These plasma properties are all closely interconnected and display highly nonlinear behavior. In addition, a limited number of actuators are available that each produce multiple different effects, making it virtually impossible to design separate controllers for each plasma property that can operate independently. Instead, control solutions need to be developed that consider the interconnectedness of the system and use the same actuators to regulate multiple different plasma properties simultaneously. In order to quantify the interconnectedness of the system, model-based control techniques can be used that rely on predictive models to describe the plasma evolution.A number of different factors determine if a predictive model is suitable for control applications. The most important requirement of these models is usually the ability to run fast enough for the relevant application; the calculation speed requirement is often on the order of milliseconds. In order to achieve these calculation speeds, many physics-based control-oriented models make simplifying assumptions, sacrificing some of their accuracy. Empirical models can achieve very high levels of accuracy at fast enough calculation speeds, but can be limited in the range of plasma scenarios they are valid for. Machine learning offers a solution to these trade-offs: by training a machine learning algorithm to replicate the calculations of a slow, high fidelity physics-oriented code, a model can be developed that runs fast enough to be useful for control applications while retaining most of the accuracy of the high fidelity code and validity across a wide range of plasma scenarios. In this dissertation, two neural network surrogate models are trained to replicate the results of physics-oriented codes: NubeamNet predicts the effects of neutral beam injection on the plasma, and MMMnet predicts the turbulent diffusivity coefficients. These neural network surrogates are integrated with conventional models to improve the fidelity of the control-oriented predictive simulation code COT-SIM. This combination of machine learning-based and conventional models are then applied to a number of different model-based control applications. A feedforward optimization scheme that uses COTSIM including neural networks as its predictive model is developed to aid in scenario planning activities. An observer algorithm is devised to estimate the state of the electron temperature profile in real time, and has been validated in real time on the DIII-D tokamak. A feedback controller is designed to actively regulate the electron temperature profile, and is shown to successfully track a temperature profile target in experiment. Another controller is developed to actively track both the electron temperature profile and the safety factor profile simultaneously.
590 $a School code: 0105.
650 4 $a Mechanical engineering. $3 649730
650 4 $a Engineering. $3 586835
650 4 $a Plasma physics. $3 3175417
653 $a Control systems
653 $a Neural network
653 $a Nuclear fusion
653 $a Tokamak
690 $a 0548
690 $a 0537
690 $a 0759
710 2 $a Lehigh University. $b Mechanical Engineering. $3 1023892
773 0 $t Dissertations Abstracts International $g 85-07B.
790 $a 0105
791 $a Ph.D.
792 $a 2024
793 $a English
856 4 0 $u https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30687905