UNM-ECE, Albuquerque, USA
SLAC, Menlo Park, California, USA
UNM-ME, Albuquerque, New Mexico, USA
ODU, Norfolk, Virginia, USA
Superconducting radio frequency (SRF) cavities with high loaded quality factors that operate in continuous wave (CW) and low beam loading are sensitive to microphonics-induced detuning. Cavity detuning can result in an increase of operational power and/or in a cavity quench. Such SRF cavities have bandwidths on the order of 10 Hz and detuning requirements can be as tight as 10 Hz. Passive methods to mitigate vibration sources and their impact in the cryomodule/cavity environment are widely used. Active resonance control techniques that use stepper motors and piezoelectric actuators to tune the cavity resonance frequency by compensating for microphonics detuning have been investigated. These control techniques could be further improved by applying Machine Learning (ML), which has shown promising results in other particle accelerator control systems. In this paper, we describe a Low-level RF (LLRF) and resonance control system based on ML methods that optimally and adaptively tunes the control parameters. We present simulations and test results obtained using a low power test bench with a cavity emulator.
UNM-ME, Albuquerque, New Mexico, USA
UNM-ECE, Albuquerque, USA
SLAC, Menlo Park, California, USA
Machine learning (ML) has recently been applied to Low-level RF (LLRF) control systems to keep the voltage and phase of Superconducting Radiofrequency (SRF) cavities stable within 0.01 degree in phase and 0.01% amplitude as constraints. Model predictive control (MPC) uses an optimization algorithm offline to minimize a cost function with constraints on the states and control input. The surrogate model optimally controls the cavities online. Time series deep ML structures including recurrent neural network (RNN) and long short-term memory (LSTM) can model the control input of MPC and dynamics of LLRF as a surrogate model. When the predicted states diverge from the measured states more than a threshold at each time step, the states’ measurements from the cavity fine-tune the surrogate model with transfer learning. MPC does the optimization offline again with the updated surrogate model, and, next, transfer learning fine-tunes the surrogate model with the new data from the optimal control inputs. The surrogate model provides us with a computationally faster and accurate modeling of MPC and LLRF, which in turn results in a more stable control system.
Machine learning, Surrogate model, control, LLRF, MPC, Transfer learning
