Project
The tokamak represents the most promising concept of fusion reactor, aimed at proving the feasibility of energy production by means of nuclear fusion. In a tokamak, a fully ionized gas of hydrogen, called plasma, is confined by magnetic fields and heated to hundred millions of degrees.
Operation of large tokamaks calls for the solution of several challenging control problems, among which a significant one is the plasma magnetic control problem, i.e. the control of the current induced into the plasma, as well as of its shape and position, by regulating the magnetic field produced by the currents flowing in external coils located around the chamber where the plasma column is formed.
Different model-based approaches have been proposed to solve the magnetic control problem in a robust fashion. As a matter of fact, algorithms with few control parameters are usually preferred, since they enable the deployment of effective adaptive algorithms, that rely on the real-time estimation of plasma parameters. Such estimation is based on the overall plasma equilibrium reconstruction, which is a computationally demanding task, that may be not compatible with the time scale the most critical magnetic control systems should react in.
Data-driven approaches represent an alternative to increase the level of robustness and adaptation capability within the available controller response time. Availability of both experiment and in-silico data does not represent an issue in the nuclear fusion domain. On the other hand, the dependance of machine learning approaches on tens or hundreds of parameters represents a twofold challenge in the field of nuclear fusion. Indeed, the performance of the control agent strongly depends on the tuning of such parameters, which is usually carried out following a trial-and-error approach. Moreover, dealing with control agents that include many parameters may represent a serious obstacle to get the license to operate the system in a fusion reactor, due to nuclear safety regulatory issues.
The Tokamak Plasma TRAINER project aims at exploiting data-driven approaches to develop fully verifiable components suitable to be deployed in a nuclear fusion power plant. The main project objectives are: i) to assess the impact of data-driven algorithm parameters on the magnetic control system performance and robustness; ii) to develop plasma parameter estimators based on data-driven identification techniques, aimed at improving the robustness of adaptive model-free controllers. A relevant by-product of the TRAINER project will be the development of a fast control-oriented nonlinear plasma equilibrium code to be used to generate in-silico data and to perform the assessment of the proposed model-free control algorithms.