Abstract
We describe a method for precisely regulating the gradient magnet power supply (GMPS) at the Fermilab Booster accelerator complex using a neural network trained via reinforcement learning. We demonstrate preliminary results by training a surrogate machine-learning model on real accelerator data to emulate the GMPS, and using this surrogate model in turn to train the neural network for its regulation task. We additionally show how the neural networks to be deployed for control purposes may be compiled to execute on field-programmable gate arrays (FPGAs), and show the first machine-learning based control algorithm implemented on an FPGA for controls at the Fermilab accelerator complex. As there are no surprise latencies on an FPGA, this capability is important for operational stability in complicated environments such as an accelerator facility.
Highlights
Particle accelerators are among the most complex engineering systems in the world
We have described a method for controlling the gradient magnet power supply (GMPS), an important subsystem of the Fermilab Booster accelerator, using machine learning models and demonstrated the feasibility of embedding such a model on a field-programmable gate array (FPGA) for a high-uptime, low-latency implementation
We trained a deep Q-network, based on a multilayer perceptron, to choose an optimal action to maximize the long-term reward, taken from the negative absolute value of the regulation error. We found this surrogate-trained network achieved a factor of 2 improvement over the existing controller in terms of the achieved rewards
Summary
Particle accelerators are among the most complex engineering systems in the world. They are crucially important to the study of the elementary constituents of matter and the forces governing their interactions. While we may model accelerator beams with impressive and improving precision [1], building fully comprehensive Monte Carlo-based models of entire facilities is challenging, often leaving optimization and control to the intuition of experienced experts Even though this process has been successful to date, it is arduous and likely contains hidden inefficiencies. We present a real-time artificial intelligence (AI) control system for precisely regulating an important subsystem of the Fermilab Booster accelerator complex [9]. In advance of commissioning the full system during online operations, we present important results from running a pretrained, static RL model in FPGA test benches Realizing this system will inevitably create an important and versatile set of tools that could find application in many areas in accelerator controls and monitoring, thereby paving the way for more ambitious control schemes.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
