Enhanced Emittance Evaluation using 2D Transverse Phase Space Distributions, High Resolution Image Denoising, and Deep Learning

TL;DR

This paper introduces an unsupervised deep learning method using a U-Net architecture to enhance the resolution and accuracy of beam emittance measurements in particle accelerators, especially under noisy and low-signal conditions.

Contribution

The work presents a novel unsupervised deep convolutional neural network framework that significantly improves image denoising and halo detection in beam diagnostics, operating efficiently on CPUs with minimal resources.

Findings

Reconstructed beam-halo structures with unprecedented resolution

Detected signals beyond seven standard deviations from the core

Revealed halo features previously unobserved at pilot sites

Abstract

Next-generation particle accelerators demand advanced beam-diagnostic capabilities to ensure high performance, operational reliability, and sustainable machine operation. Increasing beam intensities and stored energies make the precise characterization of transverse profiles, phase-space distributions, and halos often five orders of magnitude below the core but with significant environmental impact essential for loss mitigation and machine protection. Traditional analysis methods struggle with the heterogeneous, noisy, and non-Gaussian data produced under realistic operating conditions. This work presents a novel tool based on an unsupervised deep-convolutional neural-network framework that significantly enhances image denoising and restoration for emittance measurements. The method reconstructs beam-halo structures with unprecedented resolution, detecting signals at radii beyond seven standard deviations and particle densities below 10 -4 of the total beam intensity. Despite very low signal-to-noise ratios and small, non-annotated datasets, the approach preserves fine structures and reveals halo features previously unobserved at pilot installations. Built on a U-Net architecture with tailored early-stopping strategies and physics-informed metrics, the framework operates entirely on CPUs and requires minimal computational resources. The results demonstrate the potential of unsupervised deep learning as an enabling technology for high-dynamic-range beam diagnostics and motivate further development of systematic benchmarking and physics-informed learning strategies.