Particle accelerators are pushing into regimes where the difference between a smooth-running machine and a damaged one can hide in the faint outskirts of the beam. Researchers reporting in EPJ Research Infrastructures have now shown that an unsupervised deep-convolutional neural network can strip noise from beam-diagnostic images so effectively that faint halo structures, previously lost in the background, become measurable. The work, led by F. R. Osswald with M. Chahbaoui and X. Liang, targets one of the most stubborn measurement problems in accelerator physics: characterizing transverse beam profiles and halos that can sit five orders of magnitude below the core signal.
The stakes are far from abstract. Beam halos, the population of particles occupying the tails of the transverse distribution, typically represent less than 10 percent of the total beam population yet can occupy more than 70 percent of the transverse phase-space area. That disproportionate spatial footprint makes the halo the dominant contributor to beam-aperture interactions, uncontrolled losses, and the activation of accelerator components. Halo formation arises from a combination of optics mismatch, nonlinear field components, collective effects, and complex beam-environment interactions, and experimental observations show a wide variety of tail shapes that defy simple Gaussian approximations.
The operational consequences scale with machine power. The future High-Luminosity Large Hadron Collider will store 680 megajoules of beam energy, with the halo expected to contain around 35 megajoules, corresponding to a few percent of the total beam current located beyond three standard deviations of a Gaussian core. Recent operational feedback indicates that beam scraping at the HL-LHC will need to be performed at eight standard deviations rather than the five used previously, to account for transverse and longitudinal fluctuations, bunch-to-bunch variations, and turn-by-turn evolution. Similarly, the ADS project at J-PARC aims to deliver 30 megawatts of proton power, requiring scrapers positioned at fifteen standard deviations to protect quadrupole magnets from a 1 percent halo fraction. Under such conditions, beam-position errors must be controlled to keep losses below 1 watt per meter.
Measuring these faint tails demands detecting electric charges over a dynamic range spanning up to six orders of magnitude, and the fundamental challenge is separating true halo signal from noise without a reliable prior model of the noise distribution. Fluctuating experimental conditions, variable background levels, and parasitic beams all hinder universal denoising strategies. Conventional approaches fall short in familiar ways: the most common method for RMS emittance determination applies an arbitrary threshold to separate signal from background, but even modest noise levels, or the absence of minimal halo signals below 10^-4 of the total beam intensity, can introduce substantial bias, with RMS errors exceeding 100 percent in extreme cases. Dark-field subtraction, which records sensor noise in the absence of beam exposure, is also insufficient because the beam’s interaction with the scanner itself generates additional noise components that dark frames cannot capture.
The new tool was developed and tested on data from a high-resolution emittance scanner producing roughly 500 by 500 pixel grayscale images, with beams delivered from six low-energy accelerator facilities. The initial dataset comprised approximately 2,000 noisy images, but a dedicated sorting algorithm automatically rejected a substantial fraction lacking beam signal, suffering from incorrect acquisition settings, poor resolution, or severely off-center beams. Ultimately only about 10 percent of the initial dataset were retained as valid and informative samples. From this curated subset, representative images from each facility were selected by human experts, and standard analyses including fast Fourier transforms were applied to identify periodic or structured noise components.
A telling preprocessing finding shaped the entire approach: when the dataset was compared against theoretical noise models including Gaussian, Poisson, Speckle, Rician, Salt-and-Pepper, Exponential, and Gamma distributions, the resulting particle-position and angular distributions did not exactly conform to any standard model. Only by progressively raising the intensity threshold did the distributions approach Gaussian behavior, consistent with the dominance of the beam core at higher intensities. No significant correlations were detected among noise components, suggesting largely unstructured, stochastic noise. Because image properties varied so widely in size, beam shape, noise distribution, resolution, and centering, global normalization proved infeasible, reinforcing the choice of unsupervised learning in which denoising is performed on a per-image basis.
The model itself is a deep convolutional neural network with a U-shaped encoder-decoder structure and skip connections, inspired by the Deep Image Prior framework. The architecture uses two encoder blocks for down-sampling and two decoder blocks for up-sampling, two layers of 32 filters each with 3 by 3 kernels, reflection padding, bilinear up-convolutions, LeakyReLU activation, and batch normalization, with no pooling and no sigmoid at the output, allowing the network to represent real-valued physical quantities. Training relies on a custom loss function combining weighted mean squared error, mean absolute error, total variation regularization, and a gradient difference loss that preserves edge structure and texture. Notably, the analysis of learning-curve evolution shows the model exhibits what the authors describe as high impedance to noise and low impedance to signal, recovering meaningful structure early in reconstruction while resisting stochastic fluctuations.
Because the network trains on a single corrupted image with no clean target or validation set, conventional early-stopping strategies cannot be applied directly. The team instead designed adapted stopping criteria, including heuristic metrics called the Weighted Mean Variance and Edge Mean Variance, developed specifically for this architecture and for beam-diagnostic data. They validated the approach by comparing the early-stopping termination point against the optimal iteration count obtained by tracking the evolution of the beam area, derived from the RMS emittance, during extended runs of up to 100,000 iterations. When run for 10,000 iterations, the early-stopping termination point showed a strong correlation with the evolution of the RMS emittance, indicating consistency between the stopping criterion and the physically motivated optimum.
The measured results are striking. After image reconstruction, transverse amplitudes become detectable at radii beyond seven standard deviations from the beam core in the best case, and the system resolves particle populations with local densities below 10^-4 of the total beam intensity. According to the authors, this dynamic range has not previously been achieved using this class of emittance scanner. The method is also remarkably frugal: it operates entirely on CPUs, ran on a standard laptop with an Intel i5 processor and 16 gigabytes of RAM, with typical computation times below four minutes for 10^5 pixels and a few hundred iterations, well under one second per iteration, and no dependence on cloud infrastructure or internet connectivity.
The authors are careful about limitations. The present work is described as a first qualitative approach, and there remains, in their words, a long way to go before delivering a diagnostic aid tool or replacing human expertise. Deep networks can be highly sensitive to small perturbations, and even well-optimized models may produce unstable outputs under slightly altered inputs, while hidden biases and unreliable uncertainty estimates complicate trust in confidence scores. Scaling to larger image datasets raises challenges in memory requirements, convergence behavior, and heterogeneous noise management. Future directions include controlled benchmarking in which beam diameter, intensity, and resolution are incrementally varied to generate paired datasets, potentially enabling supervised models and rigorous comparative evaluations, along with physics-informed learning strategies. Even so, the demonstration that unsupervised deep learning can extract quantitative halo information from realistic, noisy, non-annotated accelerator data marks a meaningful step toward the high-dynamic-range diagnostics that next-generation machines will demand.
Subject of Research: An unsupervised deep-learning tool that denoises emittance images to improve beam diagnostics in particle accelerators.
Article Title: Deep-Learning-Enhanced Beam Diagnostics and High-Resolution Image Denoising for Next-Generation Particle Accelerators
Article References: Osswald, F. R., Chahbaoui, M., & Liang, X. (2026). Deep-Learning-Enhanced Beam Diagnostics and High-Resolution Image Denoising for Next-Generation Particle Accelerators. EPJ Research Infrastructures, 10(1), Article 21. https://doi.org/10.1007/s41781-026-00171-w
Image Credits: AI Generated
DOI: 10.1007/s41781-026-00171-w
Keywords: particle accelerators, deep learning, beam diagnostics, image denoising, U-Net, beam halo, emittance, unsupervised learning, HL-LHC, machine protection, Deep-Learning-Enhanced, Beam
News Source: Denise Maddox. (October 29, 2026). AI Cleans Noisy Accelerator Images to Reveal Hidden Beam Halo. Scienmag.
