Controlling the plasma in fusion devices is made challenging by the latency required to react to changes in the plasma state in real time as well as by the limited amount and quality of data available to the plasma control system at discharge time. But precisely these limitations make M achine L earning ( ML ) and particularly D eep L earning ( DL ) a promising alternative to traditional approaches in some workflows in plasma control. DL models differentiate themselves by the fact that most of their computational expense is paid upfront during training, where pre-existing datasets are used to tune the models’ parameters to the intended task. Training can be computationally demanding. In contrast, inferring, or making predictions, on new data using previously trained DL models is computationally trivial. This fact can be exploited to enable using complex DL models in time sensitive and data constrained real-time applications, such as plasma control. DL particularly benefits from the use of hardware accelerators, like G raphics P rocessing U nits ( GPU s), but due to the added latency associated with their use, they are still largely absent from real-time applications, limiting the benefits of using DL. In this contribution we present the technical implementation of a pipeline for the standardized and streamlined training and deployment of real-time capable DL models as augmentations to the A SDEX U p g rade ( AUG ) D ischarge C ontrol S ystem ( DCS ) Treutterer et al. (2014). A particular emphasis will be put on our simple real-time GPU-accelerated inference implementation based on commonly available tools like TensorRT NVIDIA (2019) and CUDA NVIDIA (2014). The models of our first application predict a high-fidelity electron density profile, based on the measurements of the interferometry diagnostic alone. They approximate the highly accurate I ntegrated D ata A nalysis ( IDA ) Fischer et al. (2010) profiles closely and are resilient to corruption in the input data. The complete latency of the inference process from the time of receiving data to the profile being available to the DCS is roughly 30–38 μ s running on an NVIDIA RTX L4 GPU.

In order to save flux consumption and lessen engineering requirements, the achievable toroidal loop voltage tends to be lowered in modern tokamaks. In such devices, careful optimization of the plasma startup scenario is often required to achieve a successful tokamak discharge. The PLATO tokamak, which is recently built at Kyushu University, also faces challenges in the plasma startup. PLATO employs an air-cored central solenoid that limits the loop voltage and creates stray magnetic fields. In addition, only simple capacitor banks with modest maximum charging voltages are available for the generation of the poloidal magnetic and toroidal electric fields in the initial operation. To realize the breakdown and plasma current ramp-up under those constraints, the configuration of coils and capacitor bank settings has been optimized to produce sufficient loop voltage while minimizing the stray field inside the vacuum chamber at the onset of the discharge. Through this optimization, a tokamak discharge with a plasma current of 30 kA and a pulse duration of 20 ms has been achieved. • Construction of a new air-cored tokamak that poses challenges in the plasma startup • Field coil circuits are optimized to meet the requirements of the tokamak startup • A plasma current of 30 kA and a pulse duration of 20 ms have been achieved • The total flux consumption in the discharge reaches 0.12 V ⋅ s • The electron temperature is estimated to be ∼ 67 eV

Liquid metal magnetohydrodynamic (MHD) flows in a scaled mock-up of the water-cooled lead-lithium test blanket module (TBM) were investigated experimentally in the MEKKA laboratory at the Karlsruhe Institute of Technology. The test section faithfully replicates the complex geometry of the current TBM design where liquid metal is distributed to breeder units (BUs) through long electrically coupled manifolds oriented along the poloidal direction, orthogonally to the magnetic field direction. During the experiments, pressure was recorded at various points of the mock-up for several liquid metal flow rates and magnetic fields. The overall pressure drop of the mock-up was analyzed, along with the diverse contributions to the total pressure loss from different components, based on characteristic flow parameters. A correlation for the total MHD pressure drop in the WCLL TBM was derived.

The EUropean DEMOnstration power plant (EU-DEMO) project, a EUROfusion initiative, seeks to advance fusion power technology by developing a reliable Electron Cyclotron (EC) heating and current drive system capable of delivering up to 130 MW of auxiliary heating power presently through six equatorial ports. This paper introduces an innovative alternative to the baseline ex-vessel waveguide (EW) system design, which could be major for effective EC operations. The proposed design reduces vacuum-sealed interfaces significantly — from 173 to seven — minimizing failure points and reducing the need for extensive monitoring in confined port spaces. Enhanced for remote maintainability, the design simplifies maintenance tasks and mitigates vacuum breach risks. The vacuum dimple-plated casing, acting as the primary confinement boundary, ensures nuclear safety and enables integration with current infrastructure, while addressing challenges like dynamic loads and vessel movements to maintain structural integrity. These advancements indicate that the modular and sealless waveguide approach offers a more reliable, maintainable solution. • Modular design approach ensures remote maintainability under high-radiation conditions. • Sealless waveguides cut vacuum-sealed interfaces by 96% for minimal failure points. • Dimpled-plate casing streamlines integration within the tokamak environment.