Titolo della tesi: Apply machine learning algorithms to existing experimental fusion data to advance plasma disruption and pedestal research in fusion devices
The study of tokamak disruptions is an incredibly active area for fusion researchers: when nuclear devices scale to reactor’s size and forces, the prediction and avoidance, together with mitigation of these deleterious phenomena are mandatory. As first-principles models fail to fully describe the complex non-linear interactions leading to such final loss of control, data-driven algorithms have been widely adopted to design predictive algorithms for plasma disruptions [see references 3-28 in Tinguely et al. “An application of survival analysis to disruption prediction via Random Forests” PPCF2019].
The scope of this IAEA-funded internship is to leverage existing databases of experimental data for Alcator C-Mod and TCV tokamaks to investigate stability properties in the pre-disruptive phase and therefore advance disruption research. In particular, a full statistical study of plasma equilibrium properties during the pre-disruptive phase of a specific disruption type will be conducted. Of particular interest will be an analysis of the disruption’s likelihood within a specific operational state [DeVries NF2009]: if sufficient statistics is available to describe a relevant operational space, such likelihood can provide guidance on how long the plasma can survive in a specific parameter state before a disruption occurs. Such studies have been carried out in the past for JET, with Carbon [DeVries NF2009] and ITER-like Wall [DeVries PoP2016], and similar stability maps do exist for NSTX as well [Berkery PoP2017].
A combination of classical statistical methodologies and advanced machine learning algorithms, e.g. tree-based ensemble models or deep neural networks, will be used to investigate such databases and obtain stability maps for several plasma parameters in their pre-disruptive equilibrium state. Among the plasma parameters that will be investigated: the plasma internal inductance, the profile safety factor, and the normalized Greenwald density fraction. Developing stability maps for Alcator C-Mod and TCV will likely improve our knowledge of disruption dynamics on existing devices with important consequences on our current understanding and extrapolation to future devices, such as ITER and SPARC.