Prediction of Disruptive Events on the Route to nuclear Fusion Reactors
Thursday, July 1, 2021 12:30 PM to 12:45 PM · 15 min. (Africa/Abidjan)
Stream#1
Information
Contributors:
Abstract:
Nuclear Fusion is a clean, safe, practically inexhaustible source of energy and free of greenhouse gas emissions. Large efforts are being devoted to the construction of ITER (key experimental step between today's fusion research machines and tomorrow's fusion power plants) and DEMO (demonstration fusion power plant). Both devices will be tokamaks and, therefore, the control of disruptions is of capital importance. Disruptions are sudden instabilities that develop in tens or hundreds of ms, whose effects cause a fast extinction of the plasma, with serious hazards for the first wall components. This presentation shows the evolution of methods for both the reliable prediction of disruptions, during the execution of the discharges, and the identification of the disruption types. Due to the lack of first principles and physics models to completely explain disruptive events, machine learning methods have been traditionally used for real-time prediction. Although the real-time implementations have response times of ms, the training of these models (some of them based on genetic algorithms) require a large amount of CPU time (900 h for the training of the JET APODIS predictor). With regard to off-line analysis, machine learning tools are applied to perform unsupervised clustering techniques that try to recognise in an automatic way different root causes and the physics mechanisms of disruptions. Some of the unsupervised techniques use iterative spectral clustering optimization or the framework of conformal prediction, methods that require high computational capabilities.
Abstract:
Nuclear Fusion is a clean, safe, practically inexhaustible source of energy and free of greenhouse gas emissions. Large efforts are being devoted to the construction of ITER (key experimental step between today's fusion research machines and tomorrow's fusion power plants) and DEMO (demonstration fusion power plant). Both devices will be tokamaks and, therefore, the control of disruptions is of capital importance. Disruptions are sudden instabilities that develop in tens or hundreds of ms, whose effects cause a fast extinction of the plasma, with serious hazards for the first wall components. This presentation shows the evolution of methods for both the reliable prediction of disruptions, during the execution of the discharges, and the identification of the disruption types. Due to the lack of first principles and physics models to completely explain disruptive events, machine learning methods have been traditionally used for real-time prediction. Although the real-time implementations have response times of ms, the training of these models (some of them based on genetic algorithms) require a large amount of CPU time (900 h for the training of the JET APODIS predictor). With regard to off-line analysis, machine learning tools are applied to perform unsupervised clustering techniques that try to recognise in an automatic way different root causes and the physics mechanisms of disruptions. Some of the unsupervised techniques use iterative spectral clustering optimization or the framework of conformal prediction, methods that require high computational capabilities.