Machine Learning Controls for Fermilab Accelerator Complex

The Fermi National Accelerator Laboratory FNAL/Fermilab Booster accelerator currently relies on humanintheloop monitoring of control systems to optimize beam performance during acceleration. Humans are trained to recognize undesirable beam conditions and tune away from them. This process is too slow to react to rapidly changing conditions in the Booster, and smart edge control systems are required to detect and dynamically control the beam between injection and extraction cycles 15 Hz or 66 ms. Nextgeneration control systems for the Fermi National Accelerator Laboratory Booster accelerator will use realtime edge artificial intelligence for distributed systems to control beam performance of the rapid cycling synchrotron. Machine learning inference models running in field programmable gate arrays control gradient magnet power supplies, radio frequency amplifiers and cavities, and other types of accelerator control systems that autonomously adjust gains in regulation loops as beam characteristics change in real time during operation. These models include complex deep learning and reinforcement learning models with potentially thousands to tens of thousands of weights and biases that need to be updated remotely over a local area network in real time. Crossfield’s Phase II program will extend this capability by creating a robust, endtoend framework that enables realtime updating of machine learning control parameters over a local area network. The framework will be tightly integrated into existing tools to enable rapid transition to commercial use and use inside accelerator facilities. In the Phase I program, a baseline framework was implemented that demonstrates the functionality described above. A simple clientserver model was implemented that randomly generates machine learning weights and biases on the client, streams them to a server application running under embedded Linux on the field programmable gate array, and uses direct memory access to transfer the weights and biases to the machine learning model running in the fabric. A Linux kernel device driver and a direct memory access engine were developed to enable the transfer of information and weights and biases between the server application and the machine learning core. The proposed control system framework has excellent commercialization potential in highenergy physics accelerator facilities, industrial internetofthings IIoT applications, mechatronics and robotics, and defense. The main benefit from Crossfield’s solution is the ability to automate control processes and limit human operator involvement in highspeed, largescale control systems. Modern edge artificially intelligent control systems are much more efficient and less error prone than traditional humanoperated systems.