MC6.D13 Machine Learning
SUPS032
Data-driven hysteresis compensation in the CERN SPS main magnets
use link to access more material from this paper's primary code
Magnetic hysteresis and eddy current decay continue to challenge beam quality and operational consistency in multi-cycling machines like the Super Proton Synchrotron (SPS) at CERN. Building on our previous work, this paper presents improvements in the data-driven approach for magnetic field modelling to enhance the reproducibility of SPS dipole and quadrupole fields and thus maintain stable beam parameters across all operational cycles. The method is based on feed-forward correction using magnetic field forecasting with machine learning. It now includes additional operational experience and demonstrates that the field error compensation can reliably be used in operation. This contribution proves that hysteresis compensation can be achieved without a feedback system based on expensive installations with online field measurements in reference magnets. The performance improvements achieved by eliminating the need for manual adjustments and reducing time- and energy-consuming accelerator pre-cycles are presented. The paper also sets the stage for future application in higher-order magnets, like sextupoles and octupoles, as well as on other CERN synchrotrons.
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-WEAN2
About: Received: 27 May 2025 — Revised: 04 Jun 2025 — Accepted: 04 Jun 2025 — Issue date: 05 Nov 2025
WEAN2
Data-driven hysteresis compensation in the CERN SPS main magnets
1674
Magnetic hysteresis and eddy current decay continue to challenge beam quality and operational consistency in multi-cycling machines like the Super Proton Synchrotron (SPS) at CERN. Building on our previous work, this paper presents improvements in the data-driven approach for magnetic field modelling to enhance the reproducibility of SPS dipole and quadrupole fields and thus maintain stable beam parameters across all operational cycles. The method is based on feed-forward correction using magnetic field forecasting with machine learning. It now includes additional operational experience and demonstrates that the field error compensation can reliably be used in operation. This contribution proves that hysteresis compensation can be achieved without a feedback system based on expensive installations with online field measurements in reference magnets. The performance improvements achieved by eliminating the need for manual adjustments and reducing time- and energy-consuming accelerator pre-cycles are presented. The paper also sets the stage for future application in higher-order magnets, like sextupoles and octupoles, as well as on other CERN synchrotrons.
Paper: WEAN2
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-WEAN2
About: Received: 27 May 2025 — Revised: 04 Jun 2025 — Accepted: 04 Jun 2025 — Issue date: 05 Nov 2025
THPM007
Time-varying Bayesian optimisation for continual optimal injection in the CERN PS Booster
2695
The Proton Synchrotron Booster (PSB) receives 160 MeV H- ions, which are converted to protons at injection via a charge exchange mechanism, an upgrade that allows the production of low-loss high-intensity beams (> 10^13 per ring). To mitigate losses due to space charge, horizontal phase-space painting is performed with a system of fours kickers whose pulse is customisable via time and amplitude parameters. Recent work has shown that classical optimisation algorithms can find the optimal parameter values on both a digital twin and the real machine. However, these techniques: do not handle system-state time variations, do not continually update the parameters during operation, require non-negligible dedicated beam time and are usually not robust to observation noise. We suggest time-varying Bayesian optimisation and show that it addresses each of the previous issues at low development and deployment cost. This work improves the operation of the PSB and contributes towards the goal of automating the operation of particle accelerators.
Paper: THPM007
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM007
About: Received: 19 May 2025 — Revised: 02 Jun 2025 — Accepted: 03 Jun 2025 — Issue date: 05 Nov 2025
THPM008
Virtual temperature measurements of ferrite in in-vacuum kicker magnets
2699
The Large Hadron Collider (LHC) Injector Upgrade project has achieved unprecedented beam brightness levels, to fulfill the High Luminosity LHC requirements. This higher intensity has introduced significant challenges for some of the Super Proton Synchrotron (SPS) kickers, specifically concerning beam-induced heating and vacuum rise due to electron cloud. The primary concern is the integrity of the ferrite within the kicker magnets, which is critical to the system's operation and availability. Currently, temperature monitoring relies on temperature probes (PT100) installed on the magnet's frame, but these do not provide direct measurements of ferrite temperature. To address this limitation, we present a method using deep learning techniques to develop a virtual temperature sensor, enabling real-time monitoring of ferrite temperatures across the kicker module. We apply this approach to some of the SPS injection kicker family, the so-called MKP-S, and discuss the general applicability of the method to other systems.
Paper: THPM008
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM008
About: Received: 20 May 2025 — Revised: 02 Jun 2025 — Accepted: 04 Jun 2025 — Issue date: 05 Nov 2025
THPM009
Model-based optimisation for automated multi-turn extraction tuning at the CERN Proton Synchrotron
2703
Multi-Turn Extraction (MTE) is a resonance-based technique employed in the CERN Proton Synchrotron (PS) to split the beam in horizontal phase space before extraction to the Super Proton Synchrotron (SPS). The splitting efficiency is evaluated based on the uniformity of intensities across the beamlets, requiring fine-tuning of multiple parameters. In this paper, we investigate the influence of key parameters on MTE efficiency to improve the understanding of their impact on the process. Using a Gaussian Process model and various visualization techniques, we assess the sensitivity of the MTE efficiency to horizontal tune, transverse feedback gain, excitation frequency, beam intensity and magnetic hysteresis. Results from experiments and simulations indicate a complex, non-convex relationship between MTE performance and the parameters listed above. Additionally, external factors such as thermal fluctuations may contribute to performance variability. Our findings highlight the need for a model-based controller to counteract parameter drift, thereby ensuring consistent MTE beam quality during operation. We propose a solution supported by experimental results.
Paper: THPM009
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM009
About: Received: 26 May 2025 — Revised: 30 May 2025 — Accepted: 01 Jun 2025 — Issue date: 05 Nov 2025
THPM010
Using machine learning techniques for BGI-based profile measurements at the CERN PS
2707
The Beam Gas Ionization (BGI) instrument provides a non-destructive method for monitoring transverse beam profiles by detecting free electrons produced during beam-gas ionization. Utilizing a Timepix-family detector, the BGI setup at the CERN Proton Synchrotron (PS) includes two instruments dedicated to horizontal and vertical plane measurements. However, the quality of these measurements is often compromised by artifacts, such as beam losses, which degrade profile quality, make the analysis significantly more difficult and ultimately affect the instrument performance. To address these challenges, this contribution explores the application of machine learning techniques for effective background removal. Both supervised and unsupervised approaches are evaluated on data acquired from the operational systems to improve the accuracy and reliability of the reconstructed profiles. Furthermore, the computational performance and time complexity of these methods are evaluated to ensure that the proposed solutions are compatible with the operational requirements of the BGI system.
Paper: THPM010
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM010
About: Received: 28 May 2025 — Revised: 04 Jun 2025 — Accepted: 05 Jun 2025 — Issue date: 05 Nov 2025
Machine learning-based model predictive control of the FRIB SRF
A machine learning-based model predictive control (MPC) application has been developed for the RFQ control at Facility for Rare Isotope Beams (FRIB). In this work, we extend this approach to broader applications at FRIB, the superconducting radio frequency (SRF) control. A machine learning model is trained to learn the correlations between the beam loss and the SRF signals. With the model, a MPC contoller is implemented to minimize the beam loss with high efficiency.
THPM012
Optimizing collimator positions using bayesian optimization in the Fermilab MI-8 transfer line
2711
Collimators are used to minimize losses and to remove particles that would otherwise get lost downstream and irradiate the machine. Finding the optimal jaw positions is time consuming and with the upstream beam properties changing, the collimation settings would need to be readjusted each time. Therefore, a method to optimize collimator positions and to operate them at full capacity in a short time is required for loss control downstream. A study of collimator positions was conducted and a machine learning (ML) model was developed to predict optimal collimator positions. Bayesian Optimization (BO) was used to calculate new jaw positions from the ML model. The results of BO and usage of ML for better performance of the collimation system are presented in this paper.
Paper: THPM012
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM012
About: Received: 29 May 2025 — Revised: 01 Jun 2025 — Accepted: 01 Jun 2025 — Issue date: 05 Nov 2025
Machine learning driven beam emittance optimization at EuXFEL
Planned upgrades of the European X-Ray Free Electron Laser (EuXFEL) target higher photon energy and a high duty-cycle operation up to CW operation, critically depending on improvements of the beam slice emittance of the electron gun. We are addressing this challenge by the application of deep learning techniques to create an inverse model that predicts optimal parameter configurations for the photoinjector, enabling targeted control and minimization of beam emittance. This methodology involves sampling synthetic training data through comprehensive beam dynamics simulations and introduces a machine learning-based strategy for prediction of optimal gun parameters as well as temporal pulse shaping, accommodating a family of pulse distributions including flattop and Gaussian shapes. We present results from trained neural networks with various architectures and outline our research on the invertibility of the forward model by connecting our approach to the theory of inverse problems.
Toward online learning of a cavity mechanical model for improved resonance control
The energy consumption of particle accelerators becomes an important issue nowadays. One option to address this is to employ cavities with a very high quality factor. Despite its energy saving potential, such quality factor poses a serious control problem, because the cavities become very sensitive to noise affecting their resonance frequency. A resonance controller is thus needed. There have been many attempts to design such a controller, using both model-based and model-free approaches. Yet the problem still remains an open issue. An important aspect that is apparently missing in existing solutions is a real-time adaptation to plant variations. Specifically, variations in the frequency of unwanted mechanical oscillations that perturb the cavity. In this contribution, we show the dependency of these oscillations on various operating conditions. By doing so, we motivate the adoption of a machine learning-based adaptive modeling which learns the cavity dynamics online. Such modeling is expected to improve the performance of the resonance controller by making it more robust to plant variations.
THPM016
Neural networks approach for controlling a waveform pattern of the paint bump power supply at J-PARC RCS
2715
Four horizontal paint bump magnets and two vertical paint bump magnets are used for the painting injection to produce a high intensity beam at J-PARC RCS. These paint bump power supplies are composed of the IGBT chopper units, and so the requirement waveforms can be controlled with high precision less than 1%. By using software that automatically creates input voltage (IV) patterns according to the characteristics of the power supply and direct manual adjustment of IV patterns, the current deviation of the painting pattern (PP) is less than ±0.2%.The adjustment of one PP needs about one hour and several days to adjust a total of 90 patterns with six paint bump magnets. Therefore, a reduction in adjustment time is required. In addition, to mitigate the beam loss caused by beam orbit control to a minimum level, we would like to realize a more precise current deviation. To adjust for these PP, we performed neural networks (NN) approach. By learning the IV patterns and output current patterns as training data, the highly accurate IV patterns were confirmed.The presentation reports on the preliminary training results using NN.
Paper: THPM016
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM016
About: Received: 28 May 2025 — Revised: 14 Oct 2025 — Accepted: 14 Oct 2025 — Issue date: 05 Nov 2025
THPM017
Uncertainty-Quantified Machine Model Construction Using Physics-Informed Gaussian Processes and Bayesian Optimization
2718
To construct a closed orbit model for an accelerator ring with intrinsic uncertainty quantification from orbit measurements, a physics-informed Gaussian Process model is proposed based on a stochastic ensemble of MAD-X lattices. Key advantages compared to LOCO (Linear Optics from Closed Orbits) include (1.) uncertainty-enabled orbit prediction in between BPMs (beam position monitors), (2.) fitting of a parameter distribution (dipole-like field errors) which inherently models uncertainty, (3.) incorporation of measurement uncertainty from BPM noise, and (4.) an active learning approach which can be more sample efficient than measuring an orbit response matrix. A case study is presented for the GSI heavy ion synchrotron SIS18 with various simulated applications, in particular constructing an effective machine model with minimal orbit uncertainty around the ring, and orbit correction to achieve minimal deviation at a specific location such as, e.g., the septum to control beam loss during slow extraction. This physics-inspired Gaussian Process regression approach shows potential to be applied to optics correction and further applications beyond closed orbit correction.
Paper: THPM017
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM017
About: Received: 28 May 2025 — Revised: 04 Jun 2025 — Accepted: 05 Jun 2025 — Issue date: 05 Nov 2025
THPM018
Efficient accelerator operation with artificial intelligence based optimization methods
2722
Tuning injectors is a challenging task for the operation of accelerator facilities and synchrotron light sources, particularly during the commissioning phase. Efficient tuning of the transfer line is essential for ensuring optimal beam transport and injection efficiency. This process is further complicated by challenges such as beam misalignment in quadrupole magnets, which can degrade beam quality and disrupt operations. Traditional tuning methods are often time-consuming and insufficient for addressing the complexities of high-dimensional parameter spaces. In this work, we explore the use of advanced AI methods, including Bayesian optimization, to automate and improve the tuning process. Initial results, demonstrated on the transfer line of KARA (Karlsruhe Research Accelerator) at KIT (Karlsruhe Institute of Technology), show promising improvements in beam alignment and transport efficiency, representing first steps toward more efficient and reliable accelerator operation. This study is part of the RF2.0 project, funded by the Horizon Europe program of the European Commission, which focuses on advancing energy-efficient solutions for particle accelerators.
Paper: THPM018
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM018
About: Received: 28 May 2025 — Revised: 29 May 2025 — Accepted: 30 May 2025 — Issue date: 05 Nov 2025
Machine learning-driven phase space reconstruction for heavy ion linac beams
This study explores the application of machine learning techniques for phase space reconstruction of heavy ion linac beams, a critical aspect of understanding and optimizing beam dynamics for advanced nuclear physics experiments. Modern machine learning methods, including neural networks and differentiable simulations, are employed to reconstruct the multidimensional phase space distribution from limited and noisy measurement data. These methods excel at modeling nonlinear relationships and inferring missing information, addressing traditional challenges in high-dimensional data processing. The framework uses beam diagnostics data, such as beam profiles and time-of-flight measurements, to train predictive models capable of accurately reconstructing spatial, angular, and energy distributions. Preliminary results demonstrate significant improvements in reconstruction accuracy compared to conventional approaches, with potential for real-time implementation. This work underscores the effectiveness of machine learning for beam diagnostics and optimization, offering a pathway to enhanced performance and efficiency in heavy ion linac operations.
A Bayesian multi-objective framework for optimizing an electron injector linac
This study introduces a multi-objective Bayesian optimization framework to enhance the performance of electron linear accelerators in Fourth Generation Synchrotron Radiation facilities. By focusing on minimizing horizontal and vertical emittances and energy spread at the linac exit, the approach targets improved beam quality essential for advanced synchrotron applications. Traditional methods face challenges balancing these competing objectives due to system complexity and nonlinearity. Bayesian optimization addresses this by combining probabilistic modeling and sequential sampling to efficiently navigate the high-dimensional parameter space, balancing exploration and exploitation while iteratively refining predictions. Results demonstrate the framework's ability to reduce emittances and energy spread effectively and efficiently. This scalable, adaptive method offers a robust optimization strategy for improving performance across accelerator systems in multi-objectives
THPM023
The journey towards a specialized text embedding model for accelerator physics
2730
We present PhysBERT and AccPhysBERT, specialized sentence-embedding models trained on 1.2 million arXiv physics papers and fine-tuned for accelerator physics, respectively. Evaluation across retrieval, clustering, and similarity tasks shows gains of up to 12\% over general-purpose models for physics corpora and 18\% for accelerator-specific tasks. Applications include semantic reviewer–paper matching, Retrieval-Augmented Generation for control-room logbooks, and rapid sub-domain adaptation. We analyze key design choices—data curation, masking objectives, and contrastive fine-tuning—and outline strategies for continual adaptation, providing a blueprint for domain-specific embeddings in the physical sciences.
Paper: THPM023
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM023
About: Received: 30 May 2025 — Revised: 05 Jun 2025 — Accepted: 05 Jun 2025 — Issue date: 05 Nov 2025
THPM024
Machine learning-driven longitudinal phase space reconstruction for enhanced beam tuning at LANSCE
2733
The Los Alamos Neutron Science Center (LANSCE) relies on accurate tuning of its Drift Tube Linacs (DTLs) to maintain beam quality and operational efficiency. This work introduces a novel machine-learning-based approach to reconstruct the longitudinal phase space (LPS) at the entrance of DTL Tank 1 using two-dimensional phase scans from Tanks 1 and 2. A Deep Neural Network trained on synthetic datasets generated by GPU-accelerated simulations integrates real-time diagnostic data to infer high-resolution LPS distributions. By solving this inverse problem efficiently, the method improves beam delivery precision while reducing operator intervention. Early results indicate that this approach can enhance LANSCE’s operational capabilities, providing a robust framework for accelerator tuning and diagnostics.
Paper: THPM024
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM024
About: Received: 28 May 2025 — Revised: 03 Jun 2025 — Accepted: 04 Jun 2025 — Issue date: 05 Nov 2025
THPM025
Insertion device correction based on machine learning models at the MAX IV 3 GeV ring
2737
Insertion Devices (ID) in particle accelerators introduce orbit distortions that must be compensated. At MAX IV, this compensation traditionally relies on feed-forward tables which are time-consuming to measure, and sensitive to changes in accelerator settings. This study explores the use of machine learning (ML) to automate the generation of feed-forward tables without requiring extensive measurements. Using archived data from ID gaps, beam position monitors (BPM), and corrector magnets, a neural network-based model was developed to replicate the current ID compensation system. Preliminary results show that the model effectively reproduces the existing compensation behavior and suggests potential for adaptive feed-forward tables that refine themselves with online data. In parallel, alternative ML approaches focused on minimizing beam position errors are being investigated. These efforts aim to improve the maintainability of ID compensation and prepare orbit control for future optical changes and new operational scenarios.
Paper: THPM025
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM025
About: Received: 27 May 2025 — Revised: 30 May 2025 — Accepted: 30 May 2025 — Issue date: 05 Nov 2025
THPM026
Application of Bayesian optimization for the TLS booster extraction
2740
Bayesian optimization is a method for performing global optimization on black-box functions using Gaussian processes and an acquisition function. In accelerator parameter tuning, when the number of adjustable parameters is large, finding the global optimal parameters can be time-consuming and often relies on the operator’s experience. Bayesian optimization is well-suited for such scenarios. In this report, we take the booster extraction of the Taiwan Light Source (TLS) as an example, selecting six key adjustable parameters to optimize the extraction efficiency from the booster ring to the transport line. The preliminary test results and implementation details will be discussed in this paper.
Paper: THPM026
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM026
About: Received: 08 May 2025 — Revised: 14 Oct 2025 — Accepted: 14 Oct 2025 — Issue date: 05 Nov 2025
THPM031
Beam Energy Forecasting using Machine Learning at the CLEAR accelerator
2747
In particle accelerators, accurate and stable beam parameters are crucial for experimental success. Traditional methods for measuring parameters like beam energy often rely on invasive techniques that disrupt experiments. This paper presents a novel, non-invasive machine learning-based approach to predict beam energy using parasitic measurements, enabling real-time estimation without interference. The method employs a predictive model optimized for one-step-ahead forecasting and uses time-series data decomposition to handle complex beam energy dynamics. Recursive prediction strategies allow the model to anticipate future variations autonomously. Preliminary results from experiments at the CLEAR accelerator demonstrate the model’s ability to capture both slow trends and rapid energy shifts, adapting to diverse experimental needs. These findings showcase the potential of machine learning to measure beam energy, offering a real-time, non-destructive alternative to conventional methods. This approach promises significant advancements in accelerator-based applications, especially where destructive techniques are impractical.
Paper: THPM031
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM031
About: Received: 28 May 2025 — Revised: 03 Jun 2025 — Accepted: 06 Jun 2025 — Issue date: 05 Nov 2025
THPM032
Toward autonomous control: reinforcement learning for improving CLEAR accelerator performance
2751
Particle accelerators like CLEAR (CERN Linear Accearator for research) are essential tools in advancing various scientific fields. Automating their operation to ensure stability and reproducibility is crucial for future large-scale projects. This paper explores the first steps toward autonomous control of the CLEAR beamline, focusing initially on beam steering and advancing to complex tasks like quadrupole alignment, vital for operational stability. Reinforcement Learning (RL) agents that adapt in real-time via beam screens measurements were trained and tested. The approach is optimized for sampling efficiency, addressing the high cost and invasiveness of data collection in accelerator environments. The method enables single-shot optimization for real operations, reducing the need for manual intervention. Results show that a few hours of training suffice for effective single-step corrections in the latter part of the CLEAR beamline, inspiring further development by the CLEAR research team.
Paper: THPM032
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM032
About: Received: 28 May 2025 — Revised: 03 Jun 2025 — Accepted: 05 Jun 2025 — Issue date: 05 Nov 2025
Beam orbit tuning via reinforcement learning at FELiChEM
The online optimization and debugging of particle accelerator devices have always been a challenging task. Traditional manual debugging is time-consuming and labor-intensive, and there is a phenomenon of slow drift in the machine's operating state after debugging, requiring experts to repeatedly debug. With the improvement of computing power, machine learning has developed rapidly in recent years, making it possible to train more complex models. Deep reinforcement learning models, as intelligent agents, have the ability to interact with the actual environment and learn the best strategies. By interacting with the FEL environment through deep reinforcement learning models, automatic debugging of FEL can be achieved
THPM101
Unified differentiable digital twin for the IOTA/FAST facility
2901
As the design complexity of modern accelerators grows, there is more interest in using advanced simulations that have fast execution time or produce insights about accelerator state. One notable example of additional information are gradients of physical observables with respect to design parameters produced by differentiable simulations. The IOTA/FAST facility has recently begun a program to implement and experimentally validate a unified start-to-end differentiable digital twin to serve as a virtual accelerator test stand, allowing for rapid prototyping of new software and experiments with minimal beam time costs. In this contribution we will discuss our plans and progress. Specifically, we will cover the selection and benchmarking of both physics and ML codes, the development of generic interfaces between device models and surrogate or physics-based sections, and the export of the parameters through either a deterministic event loop or a fully asynchronous EPICS soft input/output controller. We will also discuss challenges in model calibration and uncertainty quantification, as well as future plans to support larger proton accelerators like PIPII and Booster.
Paper: THPM101
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM101
About: Received: 29 May 2025 — Revised: 04 Jun 2025 — Accepted: 04 Jun 2025 — Issue date: 05 Nov 2025
THPM102
High efficiency multi-objective Bayesian algorithm for APS-U nonlinear dynamics tuning
2905
The Advanced Photon Source (APS) facility has just completed an upgrade to become one of the world’s brightest storage-ring light sources. Machine learning (ML) methods have seen extensive use during commissioning. One important application was multi-objective tuning of dynamic aperture and lifetime, a complex high-dimensionality task intractable with classic optimization methods. In this work we will discuss novel Bayesian optimization (BO) algorithmic and implementation improvements that enabled this use case. Namely, pre-training and uncertainty-aware simulation priors, dynamic parameter space and acquisition function refinement, and an adaptive wall-time convergence criteria. We will also show results of optimization runs from 10 to 24 dimensions, benchmarking scaling and efficiency as compared to standard MOGA and MGGPO. Given the promising performance, work is proceeding on tighter BO integration into the control room.
Paper: THPM102
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM102
About: Received: 29 May 2025 — Revised: 04 Jun 2025 — Accepted: 05 Jun 2025 — Issue date: 05 Nov 2025
Machine learning-enhanced infrared imaging for temperature anomaly detection in power supplies
The performance of particle accelerators is critically dependent on the reliability of their power supplies, which can number in the thousands in many facilities. In this work, we present a method for monitoring temperature anomalies in power supplies using infrared (IR) imaging. By applying various machine learning algorithms to the IR imaging data, we develop a reliable anomaly detection system that can improve the uptime of accelerator facilities. This approach enables early detection of potential issues, facilitating predictive maintenance and enhancing overall operational efficiency.
Vibration monitoring of water pumps for anomaly detection
In accelerators facilities, unexpected failures of water pumps can lead to overheating, unplanned downtime, and costly repairs. In this study, we present a novel approach for real-time monitoring of water pump vibrations to detect anomalies indicative of impending mechanical failures. We employ simple vibration sensors combined with machine learning algorithms to identify patterns and deviation from normal operating conditions. Implementation of this anomaly detection framework can significantly enhance the operational efficiency and uptime of accelerator facilities by reducing unplanned outages and extending the lifespan of water pump equipment.
THPM109
Operational results of data-driven automated intensity optimization at CERN’s LEIR
2913
At CERN’s Low Energy Ion Ring (LEIR), high beam intensities are achieved through phase space painting with up to eight multi-turn injections from the linear accelerator Linac3. After each injection, the beam is cooled and stacked in longitudinal phase space using an electron cooler. During beam operation, key parameters such as RF cavity phases in the linac, the LEIR electron cooler gun voltage, and various magnetic field strengths along the transfer line must be frequently adjusted to compensate for the injection performance degradation occurring over time. The primary cause is the aging of the stripper foil, a thin carbon plate which strips off electrons from the passing ions, altering the energy of the beam injected from the linac. Time of flight measurements in the linac and Schottky signals in the ring provide the necessary diagnostics for correcting the performance degradation and can be encoded to provide a state for an optimizer. In this paper, we compare several data-driven methods, such as Bayesian Optimization and Reinforcement Learning for designing an autonomous controller to optimize and maintain injection performance during both beam commissioning and physics runs.
Paper: THPM109
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM109
About: Received: 26 May 2025 — Revised: 02 Jun 2025 — Accepted: 04 Jun 2025 — Issue date: 05 Nov 2025
THPM110
Machine learning for the anomaly detection and characterization of the 24 GeV/c proton beam at CERN IRRAD Facility
2917
The accurate assessment of beam quality is the most important aspect in the irradiation facilities operation such as IRRAD at CERN. The Beam Profile Monitor (BPM) sensor system developed for the high-intensity proton beam at IRRAD features minimal particle interaction, improved radiation hardness and higher sensitivity and sampling rate than previous systems. It provides a wealth of high-quality BPM data not available earlier, enabling the development of data processing more advanced than before. To take advantage already today of this upgraded BPM system’s features, we propose innovative Machine Learning (ML) techniques to adapt and improve upon existing DAQ technology. This paper details the application study of (1) autoencoder architectures to perform the automatic pattern recognition and anomaly detection of proton beam profiles, and (2) deep learning techniques to predict relevant beam parameters. We applied this approach to a new dataset (made publicly available) of BPM data taken during the recent runs of IRRAD; our preliminary results demonstrate good performance in comparison to existing methods. This work is a first step towards the "intelligent" irradiation facilities.
Paper: THPM110
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM110
About: Received: 28 May 2025 — Revised: 03 Jun 2025 — Accepted: 03 Jun 2025 — Issue date: 05 Nov 2025
THPM112
Detecting anomalies in non-static environments: continual learning applied to CERN's kicker magnet
2924
The CERN accelerator complex relies critically on fast injection and extraction processes to transfer particle beams between accelerators via fast pulsed magnets, or kickers. Ensuring high availability is paramount, as the reliability of these systems directly impacts the large number of experiments conducted at CERN. In this paper, we propose to explore Continual Learning (CL) methods, specifically using Variational Autoencoders (VAEs), to develop an anomaly detection system for the fast kicker magnets. By continuously learning from evolving data while retaining prior knowledge, these models will be capable of detecting anomalies without the need for repeated retraining. This approach is particularly relevant for ensuring the reliability and stability of kicker magnets, where early anomaly detection is critical for preventing performance degradation.
Paper: THPM112
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM112
About: Received: 27 May 2025 — Revised: 04 Jun 2025 — Accepted: 05 Jun 2025 — Issue date: 05 Nov 2025
THPM113
Trajectory steering for DC beams at the CERN SPS using reinforcement learning based on intensity measurements
2928
The slow extracted beams at the CERN Super Proton Synchrotron (SPS) are transported over several 100 m long transfer lines to three targets in the CERN North Area Experimental Hall. The experiments need intensity fluctuations to be entirely eliminated over the roughly 5 s particle spill, requiring full debunching of the extracted beams. In this environment, secondary emission monitors (SEMs) have to replace the conventional beam position monitoring systems that rely on RF structure, with the intensity difference on split secondary emission foils used to indicate the beam position. Traditional trajectory correction algorithms however fail when the beam ends up on a single foil. This paper summarises successful first tests with reinforcement learning (RL) to learn to correct the trajectory based on foil intensity measurements. The RL agents were trained in simulation and then successfully transferred to the real accelerator environment. Results of the application of the trained RL agents for the alignment of moveable split foils in front of the targets will also be presented.
Paper: THPM113
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM113
About: Received: 26 May 2025 — Revised: 02 Jun 2025 — Accepted: 04 Jun 2025 — Issue date: 05 Nov 2025
THPM116
Efficient data-driven model predictive control for online accelerator tuning
2931
Reinforcement learning (RL) is a promising approach for the online control of complex, real-world systems, with recent success demonstrated in applications such as particle accelerator control. However, model-free RL algorithms often suffer from sample inefficiency, making training infeasible without access to high-fidelity simulations or extensive measurement data. This limitation poses a significant challenge for efficient real-world deployment. In this work, we explore data-driven model-predictive control (MPC) as a solution. Specifically, we employ Gaussian processes (GPs) to model the unknown transition functions in the real-world system, enabling safe exploration in the training process. We apply the GP-MPC framework to the transverse beam tuning task at the ARES accelerator, demonstrating its potential for efficient online training. This study showcases the feasibility of data-driven control strategies for accelerator applications, paving the way for more efficient and effective solutions in real-world scenarios.
Paper: THPM116
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM116
About: Received: 27 May 2025 — Revised: 02 Jun 2025 — Accepted: 04 Jun 2025 — Issue date: 05 Nov 2025
THPM120
Application of large language models for the extraction of information from particle accelerator technical documentation
2935
The large set of technical documentation of legacy accelerator systems, coupled with the retirement of experienced personnel, underscores the urgent need for efficient methods to preserve and transfer specialized knowledge. This paper explores the application of large language models (LLMs), to automate and enhance the extraction of information from particle accelerator technical documents. By exploiting LLMs, we aim to address the challenges of knowledge retention, enabling the retrieval of domain expertise embedded in legacy documentation. We present initial results of adapting LLMs to this specialized domain. Our evaluation demonstrates the effectiveness of LLMs in extracting, summarizing, and organizing knowledge, significantly reducing the risk of losing valuable insights as personnel retire. Furthermore, we discuss the limitations of current LLMs, such as interpretability and handling of rare domain-specific terms, and propose strategies for improvement. This work highlights the potential of LLMs to play a pivotal role in preserving institutional knowledge and ensuring continuity in highly specialized fields.
Paper: THPM120
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPM120
About: Received: 21 May 2025 — Revised: 04 Jun 2025 — Accepted: 04 Jun 2025 — Issue date: 05 Nov 2025
Integrated denoising for improved stabilization of RF cavities
Typical operational environments for industrial particle accelerators are less controlled than those of research accelerators. This leads to increased levels of noise in electronic systems, including radio frequency (RF) systems, which make control and optimization more difficult. This is compounded by the fact that industrial accelerators are mass-produced with less attention paid to performance optimization. However, growing demand for accelerator-based cancer treatments, imaging, and sterilization in medical and agricultural settings requires improved signal processing to take full advantage of available hardware and increase the margin of deployment for industrial systems. In order to improve the utility of RF accelerators for industrial applications we have developed methods for removing noise from RF signals and characterized these methods in a variety of contexts. Here we expand on this work by integrating denoising with pulse-to-pulse stabilization algorithms. In this poster we provide an overview of our noise reduction results and the performance of pulse-to-pulse feedback with integrated ML based denoising.
THPS072
Real-time luminosity optimization in collider experiments using reinforcement learning
3115
This study presents the development and implementation of a reinforcement learning-based algorithm for real-time luminosity tuning in collider experiments. The algorithm is initially pretrained on historical collider data and subsequently fine-tuned online during experiments. By analyzing accelerator measurements collected over several seconds, the model adjusts the magnetic structure to stabilize luminosity under varying experimental conditions. The proposed method allows for adaptive optimization without operator involvement, improving operational efficiency and stability. Results from its application on the VEPP-4M collider are presented, showcasing the method's feasibility and offering insights for its future development and application in accelerator systems.
Paper: THPS072
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPS072
About: Received: 28 May 2025 — Revised: 03 Jun 2025 — Accepted: 04 Jun 2025 — Issue date: 05 Nov 2025
THPS073
Developing an Object Detector Using Synthetic Data from CAD Models
3117
This work investigates the potential of using synthetic images generated from CAD models to train an object detector for identifying components of a particle accelerator. The study focuses on magnets within the new ALS Accumulator Ring at Lawrence Berkeley National Laboratory. Generating large volumes of real-world training data is often challenging in such complex systems. To address this, CAD files were converted into 3D models and used to produce diverse synthetic datasets. These datasets were augmented with a smaller set of real-world images to train a YOLOv8-based model. This approach aims to evaluate whether synthetic images can effectively support the development of object detectors in environments where real data collection is limited. The study lays the groundwork for future development of real-time recognition tools to assist accelerator operations.
Paper: THPS073
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPS073
About: Received: 28 May 2025 — Revised: 05 Jun 2025 — Accepted: 05 Jun 2025 — Issue date: 05 Nov 2025
Automatic online optimization at the SXFEL facility
The commissioning phase of short-wavelength FEL is often lengthy due to the optimization of thousands of control variables. These variables are frequently interdependent and have non-linear correlations with FEL performance, which makes optimization of such a complex system challenging, particularly for soft XFEL. Additionally, FEL inherently suffers from shot-to-shot intensity jitter, which necessitates online optimization in the presence of strong noise. In this study, we report the results of our experiments using an evolutionary strategy algorithm to enhance FEL intensity despite large intensity jitter.
THPS110
Advancing accelerator science through data-intensive research and training
3193
The Liverpool Centre for Doctoral Training in Innovation in Data Intensive Science (LIV.INNO) has made significant progress in applying data-intensive methods to accelerator research. This contribution presents research outcomes from the center with a focus on two key projects. The first focuses on optimizing 3D imaging for medical and industrial applications, integrating Monte Carlo simulations and advanced collimation techniques to enhance low-dose, portable X-ray systems, with implications for wider accelerator diagnostics. The second lever-ages deep learning models to reconstruct transverse beam distributions at CERN, addressing challenges in image distortion from multimode optical fibers under high-radiation conditions. The results are connected with wider progress made in machine learning and artificial intelligence for particle accelerators. Furthermore, the paper summarizes the outcomes of several key LIV.INNO events: the STFC Summer School on Data Intensive Science, the LIV.INNO 2024 Industry Showcase and the 2025 AI for Innovation Summit.
Paper: THPS110
DOI: reference for this paper: 10.18429/JACoW-IPAC2025-THPS110
About: Received: 28 May 2025 — Revised: 31 May 2025 — Accepted: 05 Jun 2025 — Issue date: 05 Nov 2025