[Paper Review] Input Beam Matching and Beam Dynamics Design Optimization of the IsoDAR RFQ using Statistical and Machine Learning Techniques

Daniel Koser, Loyd Waites+4 more|arXiv (Cornell University)|Dec 5, 2021

Particle accelerators and beam dynamics4 citations

TL;DR

This paper presents a machine learning-based surrogate modeling approach using polynomial chaos expansion and neural networks to accelerate beam dynamics optimization of the IsoDAR RFQ linear accelerator. The method enables fast virtual simulation of beam transmission and emittances, achieving high accuracy for input beam variations but facing higher errors (up to 12%) when predicting responses to structural hardware changes, making it effective for design space reduction and sensitivity analysis prior to high-fidelity simulation.

ABSTRACT

We present a novel machine learning-based approach to generate fast-executing virtual radiofrequency quadrupole (RFQ) particle accelerators using surrogate modelling. These could potentially be used as on-line feedback tools during beam commissioning and operation, and to optimize the RFQ beam dynamics design prior to construction. Since surrogate models execute orders of magnitude faster than corresponding physics beam dynamics simulations using standard tools like PARMTEQM and RFQGen, the computational complexity of the multi-objective optimization problem reduces significantly. Ultimately, this presents a computationally inexpensive and time efficient method to perform sensitivity studies and an optimization of the crucial RFQ beam output parameters like transmission and emittances. Two different methods of surrogate model creation (polynomial chaos expansion and neural networks) are discussed and the achieved model accuracy is evaluated for different study cases with gradually increasing complexity, ranging from a simple FODO cell example to the full RFQ optimization. We find that variations of the beam input Twiss parameters can be reproduced well. The prediction of the beam with respect to hardware changes, e.g. of the electrode modulation, are challenging on the other hand. We discuss possible reasons.

Motivation & Objective

• To develop a fast, accurate surrogate model of the IsoDAR RFQ beam dynamics to support online feedback during commissioning.
• To optimize RFQ design parameters for improved beam transmission and emittance control prior to construction.
• To evaluate the performance of surrogate models (PCE and neural networks) in predicting beam behavior under varying input beam and hardware parameters.
• To identify key design variables through sensitivity analysis using Sobol’s indices to reduce optimization complexity.
• To assess the limitations of surrogate models when predicting responses to structural changes in the accelerator geometry.

Proposed method

• Train surrogate models using high-fidelity PARMTEQM and RFQGen simulations as ground truth for beam dynamics.
• Apply polynomial chaos expansion (PCE) and deep neural networks (DNNs) to create fast-executing virtual RFQ models.
• Use input beam Twiss parameters and hardware design variables (e.g., electrode modulation, focusing strength) as model inputs.
• Train models on datasets of simulated beam outputs across varying input and design parameters.
• Validate model accuracy by comparing surrogate predictions against full physics simulations, especially for transmission and emittance metrics.
• Perform sensitivity analysis using Sobol’s indices to rank design variable influence on optimization objectives.

Experimental results

Research questions

• RQ1 Can surrogate models trained on beam dynamics simulations accurately predict RFQ output beam parameters such as transmission and emittances?
• RQ2 How does model accuracy vary between input beam parameter variations and structural hardware design changes?
• RQ3 What is the relative performance of polynomial chaos expansion versus deep neural networks in this context?
• RQ4 To what extent can surrogate models reduce the design space before high-fidelity optimization?
• RQ5 Why do surrogate models exhibit higher errors when predicting responses to structural changes in the RFQ geometry?

Key findings

• Neural network-based surrogate models outperform polynomial chaos expansion in predicting beam dynamics for fixed accelerator configurations, especially for input beam parameter variations.
• Surrogate models achieve high accuracy (low error) in predicting beam transmission and emittances when input beam parameters (e.g., Twiss parameters) are varied.
• When predicting responses to structural hardware changes (e.g., electrode modulation, focusing strength), surrogate model errors increase significantly, reaching up to 12% for emittances.
• The high error for structural changes persists even when space charge is turned off, indicating the issue is not due to non-linear beam dynamics but likely model generalization limitations.
• Sobol’s indices successfully identify the most influential design variables, such as DVARs 9, 10, and 13 (synchronous phase function) for transverse emittances and DVAR5 (modulation factor) for longitudinal emittance.
• Despite high errors in full hardware optimization, surrogate models are effective for preliminary design space reduction, enabling efficient follow-up with high-fidelity simulations.