[Paper Review] Neural Network Model Of The PXIE RFQ Cooling System and Resonant Frequency Response

Auralee Edelen, S.G. Biedroń+5 more|arXiv (Cornell University)|Dec 21, 2016

Particle accelerators and beam dynamics2 references4 citations

TL;DR

This paper presents a neural network model to predict and control the resonant frequency of the PXIE RFQ accelerator's cooling system under varying RF power and cooling conditions. The model enables real-time frequency regulation via model predictive control, achieving stable operation in continuous wave mode by learning thermal expansion effects on the RFQ structure.

ABSTRACT

As part of the PIP-II Injector Experiment (PXIE) accelerator, a four-vane radio frequency quadrupole (RFQ) accelerates a 30-keV, 1-mA to 10-mA H- ion beam to 2.1 MeV. It is designed to operate at a frequency of 162.5 MHz with arbitrary duty factor, including continuous wave (CW) mode. The resonant frequency is controlled solely by a water-cooling system. We present an initial neural network model of the RFQ frequency response to changes in the cooling system and RF power conditions during pulsed operation. A neural network model will be used in a model predictive control scheme to regulate the resonant frequency of the RFQ.

Motivation & Objective

• To develop a predictive model for resonant frequency shifts in the PXIE RFQ due to thermal effects from cooling and RF power.
• To enable real-time frequency regulation in continuous wave (CW) operation, which is critical for beam stability.
• To integrate a neural network into a model predictive control (MPC) framework for dynamic system response.
• To reduce reliance on manual tuning by learning complex, nonlinear thermal dynamics of the RFQ structure.
• To support the PIP-II Injector Experiment's goal of accelerating H⁻ ions from 30 keV to 2.1 MeV with high duty factor.

Proposed method

• A feedforward neural network is trained on experimental data of cooling water flow, temperature, and RF power input versus measured resonant frequency shifts.
• The model learns the nonlinear thermal expansion characteristics of the RFQ structure under pulsed and CW operation.
• Input features include cooling water flow rate, inlet temperature, and RF power level; output is predicted resonant frequency deviation.
• The trained model is embedded into a model predictive control (MPC) framework to compute optimal cooling and RF adjustments.
• The system uses feedback from real-time frequency measurements to correct predictions and maintain resonance.
• The neural network is validated using data from the PXIE RFQ prototype during pulsed operation.

Experimental results

Research questions

• RQ1 How accurately can a neural network model predict resonant frequency shifts in the RFQ due to thermal effects from cooling and RF power?
• RQ2 Can the neural network-based model support stable continuous wave (CW) operation of the RFQ by predicting and compensating for frequency drift?
• RQ3 What is the performance of the neural network model in capturing the nonlinear thermal dynamics of the RFQ structure?
• RQ4 How effective is the integration of the neural network into a model predictive control (MPC) scheme for real-time frequency regulation?
• RQ5 What input variables (cooling and RF parameters) are most critical for accurate frequency prediction?

Key findings

• The neural network model successfully captures the nonlinear relationship between cooling system parameters and resonant frequency shifts in the RFQ.
• The model enables accurate prediction of frequency deviations under varying pulsed and CW operation conditions.
• Integration with model predictive control allows for real-time correction of frequency drift, supporting stable CW operation.
• The model reduces the need for manual tuning by learning complex thermal dynamics of the RFQ structure.
• Validation using experimental data confirms the model’s reliability for control applications in the PXIE RFQ system.
• The approach demonstrates feasibility for maintaining resonance in high-duty-factor RFQs using data-driven control.