TU Darmstadt, Darmstadt, Germany
For the EPICS-based control system of the superconducting Darmstadt electron linear accelerator S-DALINAC**, supporting infrastructures based on machine learning are currently developed. The most important support for the operators is to assist the beam setup and controlling with reinforcement learning using artificial neural networks. A particle accelerator has a very large parameter space with often hidden relationships between them. Therefore neural networks are a suited instrument to use for approximating the needed value function which represents the value of a certain action in a certain state. Different neural network structures and their training with reinforcement learning are currently tested with simulations. Also there are different candidates for the reinforcement learning algorithms such as Deep-Q-Networks (DQN) or Deep-Deterministic-Policy-Gradient (DDPG). In this contribution the concept and first results will be presented.
**N. Pietralla, Nuclear Physics News, Vol. 28, No.2, 4 (2018)
TU Darmstadt, Darmstadt, Germany
The S-DALINAC* is a thrice-recirculating linear electron accelerator operating in cw-mode at a frequency of 3 GHz. A path-length adjustment system in the second recirculation beam line allows to shift the beam phase by 360° and thus to operate in ERL mode. For the multi-turn ERL operation, the beam will be accelerated twice and subsequently decelerated twice again (not demonstrated yet). For this mode, it is necessary to develop a nondestructive beam diagnostics system in order to measure the beam position, phase and beam current of both, the accelerated and the decelerated beam, simultaneously in the same beamline. A particular challenge will be the operation at low beam currents of 100 nA, which corresponds to bunch charges of about 30 aC. The conceptional study of a 6 GHz resonant cavity beam position monitor will be presented together with alternative solutions.
* N. Pietralla, Nuclear Physics News, Vol. 28, No. 2, 4 (2018)