Machine Learning Controls for Rapid Cycling Synchrotron

Fermi National Accelerator Laboratory (Fermilab), the U.S. Department of Energy, and the high-energy physics research community are interested in using machine learning to control beam dynamics in future particle accelerators, such as the Rapid Cycling Synchrotron. Fermilab requires a system that synchronously acquires data from many different types of sensors and can save the input and output data for easy access in ML frameworks. The architecture must be able to provide quick turnaround for the decision agent in reinforcement learning models locally and support ML models of the entire system. Additionally, the system requires the ability to control and configure remote devices through a graphical user interface. Crossfield Technology LLC proposes to develop a synchronous data acquisition and ML control architecture using Field Programmable Gate Array (FPGA) System-on-Chips (SoCs). The ARM processors in FPGA SoCs run embedded Linux and can control and update the FPGA fabric. Crossfield proposes to use this capability to remotely control, configure and collect data over an Ethernet network from sensors and ML algorithms running in the FPGA fabric. Crossfield plans to work with Fermilab to develop a proof-of-concept demonstration of the proposed system architecture. The Phase I will include development of a machine learning IP core for the Stratix 10 FPGA SoC and embedded Linux driver and software development to enable the demonstration. Remote users will be able to update weights in the IP remotely over an Ethernet network. The technology will be used as a testbed for future development in the Phase II program and for researchers at Fermilab. Research laboratories will benefit from a network-based machine learning control architecture that can synchronously collect data and provide remote control and configuration from a graphical user interface. The technology benefits defense and industrial applications that require similar machine learning controls in rugged environments.