Multi-objective Bayesian optimization of tokamak disruptions using fluid and kinetic models
Examensarbete för masterexamen
Complex adaptive systems (MPCAS), MSc
The generation of highly energetic runaway electron beams during tokamak disruptions is a major challenge facing tokamak fusion reactors. One of the most studied disruption mitigation schemes is massive material injection. Finding injected material densities, such that the consequences of the resulting disruption – runaway electron impact, localized heat losses and mechanical stresses – are tolerable, is still an open question, and it represents a multi-objective optimization problem. We have used a Bayesian optimization framework to optimize the injected densities of deuterium and neon in a non-activated ITER-like tokamak set up. The cost function was constructed systematically to maximize information gain, combining the maximum runaway current, final ohmic current, current quench time and conducted thermal losses. The simulations of plasma evolution were performed using the disruption modelling tool Dream. Optimization of the developed cost function was performed in two layers of physics fidelity, using both fluid and kinetic plasma models. The fluid model is computationally less expensive, which is advantageous for exploring a large parameter space. Once promising parameter regions are located using a wide search with fluid models, these are further studied in higher physics fidelity using kinetic simulations. These simulations resolve the energy distribution of the fast electrons allowing us to also account for fast electron impact ionization and energy transfer. Using two layers, the advantages of each model can be utilized resulting in an efficient optimization with a reliable examination of relevant areas. Additionally, a qualitative comparison of the two models was made to illuminate the differences between the two layers. In general, the kinetic model generated more optimistic results for the disruption consequences. More specifically, the kinetic model favoured higher neon densities and slightly lower deuterium densities compared to the fluid model. In both models, the optima are fairly insensitive to the radial distribution of neon as long as there is a higher neon density at the edge. Furthermore, the optima occurred for a moderately core-localized deuterium density. The explanation for the differences between the fluid and kinetic models was concluded to be that the fluid model overestimates the hot-tail runaway generation for certain injected material densities, resulting in larger runaway currents.
fusion plasma , disruption mitigation , runaway electron , material injection , Bayesian optimization , fluid kinetic model