Vacancy No. 35-2020-IPE

35-2020-IPE PhD Position: Embedded Deep Reinforcement Learning for the control of relativistic electron bunches on FPGA

Job description

Synchrotron light sources are used worldwide to produce brilliant light from Terahertz range to hard X-rays to investigate of a wide range of matter properties. In these sources, where electron bunches travel at relativistic velocities, a ubiquitous phenomenon occurs when the bunch charge density exceeds a certain threshold. Due to interaction between the electron bunch and its emitted electric field, microstructures spontaneously appear in the longitudinal profile (and phase space) of the bunch. In storage rings, these structures are responsible for the emission of intense coherent light in the terahertz range. Recent years have seen an increasing interest in high-power THz sources that open up a new range of research fields. Applications including characterization of superconductors, tomographic imaging, label-free genetic analysis, cellular level imaging, chemical and biological and many others. However, the dynamically changing substructures that appear during the micro-bunching instability translate into CSR power fluctuations, Figure 1 (red plot), which limits the use of the emitted THz light in experiments. During the past years the ultrafast tools for accelerator diagnostics KAPTURE and KALYPSO, have been developed at KIT. KAPTURE and KALYPSO offer MHz data-acquisition rates to enable continuous measurements on a turn-by-turn basis, and reveal the dynamics of substructures during the microbunching instabilities. Both devices contain a readout DAQ board based on novel ZYNQ UltraScale+ programmable platform designed to mean the demands of high data throughput and fast machine learning data processing.  

In this PhD thesis you will develop of a control system based on reinforcement learning (RL) for the stabilization the coherent synchrotron radiation generated by synchrotron machines. To optimize the performance of the RL controller, the hardware implementation on the ZYNQ has to been divided into two parts: the actor neural network inference, located in the FPGA, and the training policy with the critic network, located in the ARM processors. The following tasks are required:

i) design of optimal actor and critic neural networks,

ii) system integration of the RL inference within the KAPTURE and KALYPSO systems, and

iii) finally commissioning at the KIT research synchrotron KARA. 

This project requires in-depth R&D. This includes advanced design of complex FPGA firmware operating over 300 Gb/s in continuous readout mode, integration with system-on-chip for data quality check and integration to the slow-control EPICs. The project is embedded in the beam diagnostics group at the Institute of Data Processing and Electronics (IPE) at KIT. Supervision of bachelor and master students, presentations at scientific conferences, and writing high-impact journal articles is expected.

Personal qualification

A master degree in Electrical Engineering, physics or equivalent is required. Experience in development of fast neural network inference on FPGA is an advantage. You should be comfortable in specifying system components and possess sound experimental problem-solving skills. You are a naturally curious person who is eager to learn fast and has a strong interest in research. Good English language proficiency is essential, basic German language skills are of advantage.

Organizational unit

Institute for Data Processing and Electronics (IPE)

Starting date

as soon as possible

Contract duration

limited to 3 years

Application up to

31 January 2021

Contact person in line-management

For further information, please contact Dr.-Ing. Michele Caselle, phone 0721 608-25903; Mail:


Please apply online using the button below for this vacancy number 35-2020-IPE.
Personnel support is provided by 

Ms Perkovic
phone: +49 721 608-25006,

Hermann-von-Helmholtz-Platz 1, 76344 Eggenstein-Leopoldshafen, Germany

Recognized severely disabled persons will be preferred if they are equally qualified.