Evaluation of Accelerated Deep Reinforcement Learning on s ZYNC SoC and its application to electron bunch control

by Weijia Wang (KIT)

Monday Dec 14, 2020, 11:00 AM → 11:45 AM Europe/Berlin

Abstract:

 

To face up to the upcoming demand of high data throughput and fast machine learning data processing close to the data source, a readout DAQ board based on novel ZYNQ UltraScale+ programmable platform has been developed and optimized to host fast machine learning inference deployed on the FPGA.

This concept has be applied to control  Coherent synchrotron radiation (CSR) at the research synchrotron KARA. CSR is generated when the wavelength of the emitted radiation is in the order of magnitude of the bunch length. The self-interaction of short electron bunches with their own electromagnetic fields changes the longitudinal bunch dynamics significantly. Above a certain current threshold the micro-bunching instability develops, which is characterized by the appearance of distinguishable substructures in the longitudinal phase space of the bunch. During the past years, new class of world-leading tools for accelerator diagnostics: KAPTURE and KALYPSO, have been developed at IPE. KAPTURE and KALYPSO offer MHz data-acquisition rates to enable continuous measurements on a turn-by-turn basis, which was previously impossible, therefore revealing the dynamics of substructures during the microbunching instability.

To stabilize the CSR emission, a real-time feedback control loop based on Reinforcement Learning (RL) is proposed.  To satisfy the low-latency requirement, the RL network has been implemented on hardware (FPGA).  In this talk, the real-time feedback loop architecture and its performance will be presented and compared with Keras-RL on CPU/GPU.

 

Language: English

Towards Intelligent Data Acquisition Systems with Embedded Deep Learning on MPSoC.pdf