Introduction ------------ The physicist Andreevich Artsimovich in the 1970 wrote that "thermonuclear [fusion] energy will be ready when mankind needs it". Considering the actual world energy situation and the effect on the environment due to the present harnessing of the different sources of energy, the hope is that time for fusion is finally arrived. Background and Motivation ------------------------- The activities carried out in the framework of this thesis regarded the devel- opment, implementation and application of algorithms for classification and prediction of disruptions in Tokamaks. The balance of plasmas in a magnetic field can be described by the theory of magneto-hydro-dynamic (MHD). MHD instabilities are among the most serious factors that limit fusion devices operation in magnetic confinement configurations. When they occur on a large scale can degrade the perfor- mance of the plasma and lead to loss of confinement and control. A disruption is a sudden loss of stability or confinement of tokamak plasma; it is a critical event in which the plasma energy is lost within a time span of few milliseconds exposing the plasma facing components to se- vere thermo-mechanical stresses and conductors surrounding the vessel to huge electromagnetic forces. Therefore, it becomes of primary importance to avoid or mitigate disruptions in order to preserve the integrity of the ma- chine. This aspect and the understanding of disruptive phenomena play a key role in design and running of new experimental devices as ITER, cur- rently under construction in Cadarache (France), which will have the task of demonstrating the feasibility of fusion energy production from a technical and engineering point of view. These considerations motivate a strong interest in developing methods and techniques aimed to minimize both number and severity of disruptions. Furthermore when a disruption occurs it would be particularly important to be able to distinguish among its difierent types in order to improve avoidance and mitigation strategies. Since physical models able to reliably recognize and predict the occurrence of disruptions are currently not available, the re- search carried out fits in the broad framework of machine learning techniques that have been exploited as an alternative approach to disruption prediction and automatic classification. Promising approaches to prediction and classification are represented by the so-called "data-based" methods: to this purpose, existing systems have been applied and further developed and new approaches have been investi- gated. The mentioned activity has been carried out in collaboration with the University of Cagliari and European Research Centers for nuclear fusion, taking as case study some of the most important experimental machines such as JET and ASDEX Upgrade (AUG), with several months of research spent at the Culham Science Centre. Outline of the Thesis --------------------- In chapter 1 the perspectives of fusion in the world energy context as an almost unlimited source of energy for the future are discussed, with particu- lar reference to the role of magnetic confinement. Furthermore, the bases of fusion reactions have been introduced. In chapter 2 the main aspects of plasma stability in tokamaks configu- rations are described with the aim to provide an adequate reference for all the discussions of the following chapters. In particular, the main parameters related to plasma stability, which have been used for the construction of the databases, have been introduced. The chapter 3 is focused on the description of the operational limits with reference to the main quantities which should be maximized to im- prove plasma performance. Everything, also in the previous chapters, has been framed to introduce the key problems which this thesis has addressed: analysis, prediction and classification of disruptions. After the main consid- erations about the operational limits, the main phases, the causes and the consequences of disruptions have been discussed, trying to integrate the sta- bility concepts introduced in the previous chapter. The chapter 4 is finalized to provide an insight of the Machine Learn- ing methods which represent the starting point of all the analysis and algo- rithms implemented for disruption prediction and classification. Today the large amount of data available from fusion experiments and their character of high-dimensionality make particularly difficult handling, processing, un- derstanding and extracting properly what is really important among all the available information. Machine Learning allows to deal with the problem in efficient way. Therefore, a framework of all the techniques exploited for the analysis has been provided, with particular reference to the Manifold Learn- ing algorithms as Self Organizing Maps (SOMs) and Generative Topographic Mappings (GTMs). Also reference methods such as k-Nearest Neighbor (k- NN) or more recent methods such as Conformal Predictors, exploited for validation and reliability assessment purposes, have been described. In chapter 5 the state of the art of machine learning techniques ap- plied to disruption prediction and classification is presented, describing in particular the main applications with the widely employed Neural Networks, such Multi Layer Perceptrons (MLPs), Support Vector Machines (SVMs) and Self Organizing Maps (SOMs), and statistical methods such as Discrim- inant Analysis or Multiple Threshold technique. Strengths and weaknesses have also been discussed with reference to a possible solution to overcome the drawbacks of these methods: the multi-machine approach. Chapter 6 is dedicated to the description of the databases used for all the analysis presented in the following chapters. In particular, the statistical analysis and the data-reduction algorithms that have been needed to build a reliable and statistically representative database have been discussed in detail. The last three chapters contain all the analysis and all the algorithms im- plemented for the mapping of the operational space, disruption classification and prediction. In chapter 7 the mapping of the JET operational space is presented. The first sections deal with projections and data-visualization with linear projection methods such as Grand Tour (GT) and Principal Com- ponent Analysis (PCA). In the central part, the same aspects have been taken into account by exploiting nonlinear Manifold Learning techniques, SOM and GTM, on the base of which a detailed analysis of the operational space has been performed. Such analysis, showing the potentiality of the methods, has been performed, regarding GTM model, through the implementation of a dedicated tool. Finally, an outliers' analysis and performance indexes appo- sitely proposed have been considered for evaluating the overall performance of the mapping. In the chapter 8 the developed automatic disruption classification for JET has been described. The chapter is divided in two parts: the first one describes the classification of disruptions belonging to the Carbon Wall (CW) campaigns, whereas in the second part the classification of disruptions with the ITER-like Wall (ILW) is framed in the assessment of the suitability of the automatic classifier for real time applications, in conjunction with prediction systems working online at JET. The reliability of the results has been vali- dated by comparison with a k-NN based reference classifier and through the recent conformal predictors, with which is possible to provide, in addition to the prediction/classification, the related level of confidence. Chapter 9 is dedicated to the disruption prediction at ASDEX Upgrade. The first part is related to the description of the database and the data- reduction technique used to select a representative and balanced dataset. Self-Organizing Map and the Generative Topographic Mapping have been exploited to map ASDEX Upgrade operational space and to build a disrup- tion predictor, introducing at the same time their potentiality for disruptions classification. Furthermore, the use of this two methods combined with a Lo- gistic model has been proposed to realize a predictive system able to exploit the complementary behaviors of the two approaches, improving the overall performance in prediction.