5th ICFA Beam Dynamics Mini-Workshop on Machine Learning for Particle Accelerators

8 Apr 2025, 08:00 → 11 Apr 2025, 19:30 Europe/Zurich

80/1-001 - Globe of Science and Innovation - 1st Floor (CERN)

Francesco Maria Velotti (CERN), Verena Kain (CERN)

5th ICFA Beam Dynamics Mini-Workshop on Machine Learning for Particle Accelerators

 

We are pleased to announce the 5th ICFA Beam Dynamics Mini-Workshop on Machine Learning for Particle Accelerators. 

The goal of this workshop is to bring together the world-wide community of researchers applying machine learning techniques to particle accelerators.

Do not hesitate to share the news (https://indico.cern.ch/e/icfa-ml-2025) with your colleagues and contact the organizers (icfa-ml-2025@cern.ch) with questions or proposals for contributions.

The workshop will consist of these topics:

• Surrogate modelling/digital twins
• Optimisation and Control
• MLOps, Infrastructure, Scalability
• Anomaly Detection and Diagnostics
• LLMs/AI assistants

 

Tutorials and lecture topics:

• Introduction to Control Theory
• Differentiable simulations
• Multi-objective RL
• LLMs for knowledge retrieval - an introduction

 

 

Tuesday 8 April

08:30 → 09:00 Registration

09:00 → 09:20 Introduction and welcome

Speakers: Tia Miceli, Verena Kain (CERN)

MaLAPA 2025 Discussion Sessions.pdf

Welcome.pdf

09:20 → 09:50 Highlights from RL4AA'25

Speakers: Andrea Santamaria Garcia (University of Liverpool), Annika Eichler (DESY)

RL4AA25 MaLAPA Recap.pdf

09:50 → 10:45 Introduction to Control Theory

Control theory is a pivotal field of study that focuses on the behavior of dynamical systems and the development of strategies to influence these systems towards desired outcomes. The principle of control theory find its application in plenty of disciplines including engineering, economics, biology and beyond. It were control concepts like the Kalman filter that has flew the Apollo to the moon. Also for the operation of particle accelerator there are thousands of control loops running, ranging from low-level hardware-focused systems, as magnet powering, synchronization or RF control, over fast transverse feedback systems to beam steering.
In this tutorial, we begin by exploring the basic concepts of control, emphasizing the importance of stability and robustness. The tutorial then introduces the mathematical modeling of systems using differential equations and the representation of these systems in state-space form. Starting from classical control strategies, we will go towards modern optimal control techniques and will finally see the link to optimization and ML approaches like reinforcement learning.

Speaker: Annika Eichler (DESY)

Tutorial_Control.pdf

10:45 → 11:15 Coffee Break

11:15 → 12:15 Multi-objective RL

Multi-objective reinforcement learning (MORL) extends traditional reinforcement learning (RL) by addressing environments where multiple conflicting objectives must be optimized simultaneously. In real-world applications, such as autonomous systems, particle accelerator optimization and control, agents often face trade-offs between competing goals. This lecture provides an overview of the key concepts, techniques, and challenges in MORL. We will explore different approaches to handling multiple objectives, including scalarization methods, Pareto dominance, and reward decomposition. Additionally, we will discuss the concept of the Pareto front and how it can guide decision-making in complex environments. The lecture will also highlight the role of exploration-exploitation trade-offs in multi-objective settings and the development of algorithms that can balance these objectives effectively. By the end of the session, attendees will gain a deeper understanding of how to design and implement multi-objective RL systems capable of navigating the complexities of real-world decision-making tasks.

Speaker: Kishan Rajput (Jefferson Lab)

Kishan_MaLAPA_5_Lecture_MORL.pdf

Kishan_MaLAPA_5_Lecture_MORL.pptx

12:15 → 14:00 Lunch

14:00 → 15:00 Differentiable simulations

Many accelerator physics problems, such as beamline design, beam dynamics model calibration, online tuning and phase space measurements rely on solving high-dimensional optimisation problems over beam dynamics simulations. Numerical optimisers have successfully been applied to such tasks, but they struggle as the dimensionality and complexity of the objective function increase. In machine learning, gradient-based optimisation algorithms are successfully used to optimise billions of model parameters over complex loss functions when training large neural network models. This is made possible by reverse-mode automatic differentiation, which enables the fast computation of gradients of complex functions. In this tutorial, you will learn to use novel beam dynamics simulations with support for automatic differentiation to your advantage and harness the power of gradient-based optimisation in accelerator physics. Multiple hands-on examples using the Cheetah beam dynamics code will allow you to try these methods for yourself. While we will present multiple example applications of gradient-based optimisation on differentiable beam dynamics simulators, the space of potential applications here is vast, and we believe that participants will go on to discover numerous novel applications for differentiable beam dynamics simulations that were intractable to solve with existing methods.

Speakers: Andrea Santamaria Garcia (University of Liverpool), Chenran Xu, Jan Kaiser, Juan Pablo Gonzalez Aguilera (University of Chicago)

Tutorial GitHub Repository

Tutorial Slides

15:00 → 15:30 Coffe break

15:30 → 17:10 Optimisation and Control

Convener: Michael Schenk (CERN)

15:30 Serval Applications of Machine Learning at HEPS- 15'+5'

The High Energy Photon Source (HEPS) is a fourth-generation synchrotron radiation facility under construction in Beijing, China. Since the beam commissioning of the storage ring commenced in July 2024, progress has proceeded smoothly, and the first light was achieved in October. During the construction and beam commissioning of HEPS, we explored machine learning to address critical technical challenges. In the offline optimization phase, we applied serval machine learning methods, focusing on data-driven evaluation and optimization of Touschek lifetime. With approximately 95% evaluation accuracy, we reduced computation time by about 90% compared to traditional methods, thereby significantly improving computational efficiency. In the beam commissioning phase, we employed unsupervised learning methods to identify abnormal states in the power supplies, successfully detecting anomalies. Furthermore, we are optimizing beam lifetime by applying machine learning algorithms to accelerator parameters and Beam Loss Monitor (BLM) data in an ongoing study. In recognition of the importance of high-quality data for machine learning applications, we developed the FAIR-Compliant AI-Ready Accelerator Data Platform (FARAD), aimed at generating AI-ready datasets to advance research in the accelerator field. All the beam experiments described above were conducted on the FARAD platform, greatly enhancing research efficiency. This paper will provide a detailed description of the progress in the research outlined above, along with future plans.

Speaker: wei bao

Serval ML applications at HEPS.pptx

15:50 Photon systems automation activities at EuXFEL 15'+5'

Large-scale facilities like European XFEL consist of a multitude of subsystems, which often require frequent calibration. Additionally, accurate real-time tuning of many of these subsystems is critical to maintain stable and optimal performance. Automation techniques can be leveraged to reduce operators' time investment and potentially increase the exploitation of allotted beamtime, both in quantity and quality. One example is the automation of photon beam alignment through an instrument, achieved by adjusting multiple optical components, such as mirrors and lenses, using precise and constrained actuation. This process uses Bayesian optimization to iteratively determine the optimal configuration by evaluating system performance metrics, such as beam intensity, shape and position, which requires reliable image processing. In this talk, I will introduce ongoing activities aimed at automating selected components of the photon system at the European XFEL.

Speaker: Sarlota Birnsteinova (European XFEL GmbH)

ICFAworkshop2025_Birnsteinova.pdf

16:10 ML optimization methods for APS-U commissioning- 15'+5'

The Advanced Photon Source (APS) facility has just completed an upgrade to become one of the world’s brightest storage-ring light sources. For the first time, machine learning (ML) methods have been extensive used as part of the baseline commissioning plan. Most popular such method was Bayesian optimization (BO) – a tool for efficient online high-dimensional single and multi-objective tuning. In this paper we will present our BO development work on experimentally motivated augmentations - uncertainty-aware simulation priors, parameter space and acquisition function refinement for multi-objective optimization, and online execution time improvements. These improvements were integrated into the APSopt optimizer, which was then successfully used for various commissioning tasks. We will show results of tuning linac and booster transmission efficiency, injection trajectory stabilization, and of extensive multi-objective storage ring dynamic/momentum aperture studies. Given the success of BO methods, work is proceeding on tighter integration into the control room.

Speaker: Nikita Kuklev (Fermilab)

icfaml25_nk_mobo_talk.pdf

16:30 Data-Driven Feedback Optimization for Particle Accelerator Control- 15'+5'

Feedback control is an essential component for the successful operation of particle accelerators. However, achieving the desired closed-loop performance requires precise model knowledge, which is difficult to obtain in complex accelerator systems. For this reason, we present an application of a combined optimization approach that estimates the response matrix online while optimizing the chosen performance measure, eliminating the need for first-principles modelling or a priori identification experiments. Specifically, a Kalman filter is employed to construct a linearization of the system response around its operating point from noisy input-output measurements, iteratively improving the available knowledge about the system. In parallel, this knowledge is exploited by a feedback optimizer, which is incrementally driving the system to its optimal operating point while maintaining safety of operation as formulated by constraints on inputs and outputs. As a consequence of the continuous online response estimation and in contrast to other modelling approaches, the scheme is able to instantaneously react to changes and drifts in the system behaviour. This is demonstrated on the orbit feedback of the main electron beam dump line of the European XFEL.

Speaker: Christian Hespe (DESY)

2025_04 MaLAPA Feedback Optimization.pdf

2025_04 MaLAPA Feedback Optimization.ppt

16:50 Machine Learning for Online Control of Particle Accelerators- 15'+5'

Particle accelerators play a critical role in modern scientific research. However, existing manual beam control methods heavily rely on experienced operators, leading to significant time consumption and potential challenges in managing next-generation accelerators characterized by higher beam current and stronger nonlinear properties. In this paper, we establish a dynamical foundation for designing the online adaptive controller of accelerators using machine learning. This provides a guarantee for dynamic controllability for a class of scientific instruments whose dynamics are described by spatial-temporal equations of motion but only part variables along the instruments under steady states are available. The necessity of using historical time series of beam diagnostic data is emphasised. Key strategies involve also employing a well-established virtual beamline of accelerators, by which various beam calibration scenarios that actual accelerators may encounter are produced. Then the reinforcement learning algorithm is adopted to train the controller with the interaction to the virtual beamline. Finally, the controller is seamlessly transition to real ion accelerators, enabling efficient online adaptive control and maintenance. Notably, the controller demonstrates significant robustness, effectively managing beams with diverse charge mass ratios without requiring retraining.This controller enables global control, achieving up to 42-dimensional synchronous regulation across the entire superconducting section of the China Accelerator Facility for Superheavy Elements.

Speaker: Xiaolong Chen (Institute of modern physics)

ML for OnlineControl-v1.pptx

17:10 → 18:10 Community session: MALAPA Community Resources

Conveners: Jan Kaiser, Ryan Roussel

20250410_ICFA_AIML_Roadmap_Discussion.pdf

20250410_ICFA_AIML_Roadmap_Discussion.pptx

18:45 → 20:45 Cocktails with Jazz and finger food

Wednesday 9 April

08:30 → 10:30 Optimisation and Control

Convener: Seongyeol Kim

08:30 Physics-informed Bayesian inference and optimization of the closed orbit in synchrotrons -15'+5'

Heavy ion synchrotrons, like the SIS18 at GSI, rely on the proven numerical approaches to correct the closed orbit. The SIS18 has a relative moderate amount of BPMs (one per cell) and requires a well corrected and known orbit, especially near the injection/extraction systems. Fluctuations of the BPM signal arise from the electronics. In addition there are systematic errors due to the relative positioning of the BPMs. At specific locations, like the the beam position at the extraction septum, it is desirable to have a prediction, including an uncertainty estimate. An adapted approach towards closed orbit correction is proposed that integrates probabilistic modeling with beam dynamics to infer a closed orbit including uncertainty quantification. Methods, such as LOCO (Linear Optics from Closed Orbits) and NOECO (Nonlinear Optics from Off-Energy Closed Orbits), are limited by the need for extensive orbit response matrix (ORM) measurements and lack uncertainty quantification. The proposed method leverages physics-informed Bayesian regression to develop a surrogate model that not only quantifies uncertainties at beam position monitors (BPMs) but also in between them, reducing the required data. A Gaussian Process (GP) model is used to incorporate beam dynamics by estimating the kernel (and mean function) through the evaluation of simulated realizations, with simulations based on a MAD-X model of the SIS18 lattice. The learned distribution of multipole misalignments enables a model of the closed orbit with integrated uncertainty and noise handling. This model is then used in a Bayesian optimization framework to correct the closed orbit and achieve minimal deviation at specific locations, such as at the septum. The approach has also broader applications towards more general optics corrections.

Speaker: Victoria Isensee

ICFA_ML_2025_Presentation_VIsensee.pdf

08:50 Optimising Injection Efficiency at Diamond Light Source using Gaussian Processes with Non-Gaussian Likelihoods - 15'+5'

A general Bayesian optimisation tool is being developed at Diamond Light Source to improve machine performance by constructing surrogates from Gaussian Process (GP) models. Priors are placed on covariance kernel hyperparameters to guide an optimiser and prevent overfitting. The model has been integrated with the machine control system. During an experiment aimed at improving injection efficiency by changing magnet currents, hysteresis was identified and compensated. Multiple measurements were taken at each current to determine the measurement error. However, large variations in the error with respect to current were observed; this heteroscedastic behaviour was handled robustly with non-Gaussian likelihoods incorporated into the inference step. The injection efficiency was increased running the model in less time than a manual scan from operators.

Speaker: Shaun Preston (John Adams Institute, University of Oxford)

ICFA-2025-S-Preston

09:10 Eliminating mains noise with Machine Learning- 15'+5'

Power supply ripples at various frequencies - characteristic to the magnet circuits or from the electrical network - have always been an issue in accelerator operations, with several mitigation measures put in place over the years. This contribution summarises recent efforts in the CERN SPS to compensate the ripple at 50 Hz and its harmonics in the main quadrupole circuits with ML methods. It will start with introducing the detrimental effects of the ripple at low energy for LHC-type beams and at top energy for slow extracted beams. For optimal conditions of slow extracted beams, a continuous control algorithm had to be conceived. The implementation required hardware modifications on the power converter electronics side, additional new controls infrastructure and the development of adaptive algorithms that can deal with the changes of the electrical distribution network throughout the day. Adaptive continuous control with adaptive Bayesian Optimisation has been in place for slow extracted spill control throughout 2024. The obtained improved spill quality over the year will be discussed. First impressive results with 50 Hz compensation for the LHC ion cycles in the SPS during the ion run at the end of 2024 will also be presented. Finally, ideas for further R&D towards one-shot corrections for beams that are only played on-demand (i.e. LHC beams) will be proposed.

Speaker: Verena Kain (CERN)

MaLAPA-50Hz_2025.pdf

09:30 Beam halo losses reduction with simulation constrained Bayesian Optimization- 15'+5'

The Linear IFMIF Prototype Accelerator (LIPAc) is designed to accelerate 125 mA of D+ to 9 MeV in CW. The very high power stored in the beam (~1.1 MW) and the use of superconductive RF cavities requires precise control of beam losses (target <1e-6). On the other hand the intense beam is affected by strong space charge forces that easily results in significant halo formation. This contribute is difficult to simulate because requires large number of particle tracking and precise knowledge of input phase space. In this work we present the use of Bayesian optimization of transport optic (4 quadrupoles + 4 steerers) to minimize the halo losses by observing vacuum response in the sector. As a single beam pulse with inappropriate optics could lead to permanent damage of the machine, we include constraint of negligible losses from beam core in simulations. The algorithm successfully and safely reduced vacuum in sector by a factor of ~3 within two hours of operation. After analysis of newly set-point we find that the proposed optics results both in a compromise matching of core and halo distribution and a reduction of particles ejection from the former to the latter.

Speaker: Andrea De Franco (National Institutes for Quantum Science and Technology (QST))

2025_04_08_ML_Halo_DeFranco.pdf

2025_04_08_ML_Halo_DeFranco.pptx

09:50 Efficient Dynamic and Momentum Aperture Optimization for Lattice Design Using Multipoint Bayesian Algorithmic Execution- 15'+5'

This paper presents a novel application of Multipoint Bayesian Algorithmic Execution (multipointBAX) to optimize dynamic aperture (DA) and momentum aperture (MA) in lattice design. DAMA optimization is a critical design task for storage rings, ultimately determining the flux of x-ray sources and luminosity of colliders.  Traditionally, solving this multi-objective optimization problem has relied on genetic algorithms (GA) and/or Bayesian optimization (BO), requiring extensive particle tracking simulations for each trial configuration, which in turn limits the quality of the final design. Here we instead use multipointBAX to select, simulate, and model a single particle for each trial configuration, resulting in two orders of magnitude higher end-to-end efficiency, while also integrating random magnet error seeds to increase robustness. The multipointBAX implementation involved several advances, including a neural-network surrogate model, a batch acquisition strategy, and the concept of a "Pareto front region" to improve stability. A proof-of-principle demonstration on the SPEAR3 storage ring design validates our approach, with multipointBAX achieving equivalent Pareto front results with only 1% of the tracking computations required by GA.  We are now applying the method to the design of future light sources and colliders.

Speakers: Daniel Ratner (SLAC), Zhe Zhang

DAMA_Apr2025.pdf

DAMA_Apr2025.pptx

10:10 Progress on Automating Experiments at the Argonne Wakefield Accelerator Using Advanced Bayesian Optimization Algorithms - 15'+5'

The goal of machine learning for accelerator control is to automate the start-up, optimization, and execution of experiments at accelerator facilities with limited-to-no human operator input. To address this challenge, we have been pursuing a research program to completely automate sequential accelerator beamline configuration tasks at the Argonne Wakefield Accelerator (AWA). These tasks include beam alignment with magnetic elements, RF cavity phasing synchronized to the photoinjector laser system, and transverse phase space control, all in tightly constrained parameter spaces. In this work, we describe progress towards automating the AWA beamline and the development of novel Bayesian optimization algorithms needed to efficiently address each of these tuning tasks. We highlight the development of Bayesian Algorithm Execution (BAX) acquisition functions to generalize Bayesian optimization algorithms to optimize so-called “virtual” objectives that are inferred from Gaussian process models of observable quantities. Finally, we describe the path forward for achieving full autonomous execution of beamline configuration and experiment execution in the near future.

Speaker: Ryan Roussel

rroussel_MALAPA_2025.pptx

10:30 → 11:00 Coffee break

11:00 → 12:00 Optimisation and Control

Convener: Jason St John (Fermi National Accelerator Laboratory)

11:00 Harnessing the Power of Gradient-Based Simulations for Multi-Objective Optimization in Particle Accelerators- 15'+5'

Particle accelerator operation requires simultaneous optimization of
multiple objectives. Multi-Objective Optimization (MOO) is particularly challenging
due to trade-offs between the objectives. Evolutionary algorithms, such as genetic
algorithm (GA), have been leveraged for many optimization problems, however, they
do not apply to complex control problems by design. This paper demonstrates
the power of differentiability for solving MOO problems using a Deep Differentiable
Reinforcement Learning (DDRL) algorithm in particle accelerators. We compare
DDRL algorithm with Model Free Reinforcement Learning (MFRL), GA and Bayesian
Optimization (BO) for simultaneous optimization of heat load and trip rates in the
Continuous Electron Beam Accelerator Facility (CEBAF). The underlying problem
enforces strict constraints on both individual states and actions as well as cumulative
(global) constraint for energy requirements of the beam. A physics-based surrogate
model based on real data is developed. This surrogate model is differentiable and allows
back-propagation of gradients. The results are evaluated in the form of a Pareto-front
for two objectives. We show that the DDRL outperforms MFRL, BO, and GA on high
dimensional problems.

Speaker: Kishansingh Rajput (Jefferson Lab)

GradRL_talk.pptx

11:20 Virtual to Physical: Reinforcement Learning to Optimize SNS Particle Accelerator Controls- 15'+5'

Complex accelerators must have control systems that can handle dynamic nonlinear environments. This makes traditional control methods unsuitable as they can struggle to adapt to these uncertainties. This provides an ideal environment for reinforcement learning algorithms as they are adaptable and generalizable. We present a reinforcement learning pipeline that can effectively handle the dynamics of a complex accelerator. We test and prove our pipelines capabilities on multiple environments including the Spallation Neutron Source (SNS) and the Beam Test Facility (BTF) at Oakridge National Lab (ORNL). Due to the limited time available to train an online algorithm like reinforcement learning on a real accelerator, we utilize a virtual twin accelerator (VIRAC) developed by ORNL to pretrain the policy and show its ability to converge in the virtual environment. We then test the adaptability of the pretrained RL model by applying it on the real accelerator and comparing the results. Utilizing our Scientific Optimization and Controls Toolkit (SOCT) and open-source standards such as Gymnasium we create and solve for a MEBT orbit correction problem in the SNS and an emittance maximization problem in the BTF. We show how Twin Delayed Deep Deterministic Policy Gradient (TD3) can solve this optimization environment in the virtual accelerator and transfer this policy onto the real accelerator for inference and model retraining. We show how reinforcement learning can be utilized as a control system for complex accelerators and provide a model pipeline for how an implementation performs and can be adapted to new accelerator control problems.

Speaker: Armen Kasparian (Jefferson Lab)

RL4SNS - ACK.pdf

RL4SNS - ACK.pptx

11:40 Explainable physics-based constraints on reinforcement learning for accelerator controls- 15'+5'

We present a reinforcement learning (RL) framework for controlling particle accelerated experiments that builds explainable physics-based constraints on agent behavior. The goal is to increase transparency and trust by letting users verify that the agent's decision-making process incorporates suitable physics. Our algorithm uses a learnable surrogate function for physical observables, such as energy, and uses them to fine-tune how actions are chosen. This surrogate can be represented by a neural network or by a sparse dictionary model. We test our algorithm on a range of particle accelerator controls environments designed to emulate the Continuous Electron Beam Accelerator Facility (CEBAF) at Jefferson Lab. By examining the mathematical form of the learned constraint function, we are able to confirm the agent has learned to use the established physics of each environment. In addition, we find that the introduction of a physics-based surrogate enables our reinforcement learning algorithms to reliably converge for difficult high-dimensional accelerator controls environments.

Speaker: Jonathan Colen (Old Dominion University)

ExplainableConstraints.pdf

12:00 → 12:45 MLOps, Infrastructure and Scalability

Convener: Andrea Santamaria Garcia (University of Liverpool)

12:00 Machine learning and computer vision in robotic interventions at CERN boosting maintainability- 10'+5'

Robots are used in the CERN accelerator complex for remote inspections, repairs, maintenance, monitoring, autopsy and quality assurance, to both improve safety and machine availability. Past interventions mostly relied on teleoperation of robotic bases and arms, while some current and many future interventions will use autonomous behaviors, largely based on advances in machine learning and computer vision. This transition will reduce the operator workload and allow less expert operators to use robots. Autonomous execution can also ensure consistent control and measurement conditions, which are crucial for maintaining repeatability of results, particularly for radiation surveys which data can be used in historical comparisons, and sensor calibrations. As an example, mobile robots in the SPS currently use an advanced autopiloted measurement procedure with adaptable parameters, a dual solution for automated safety gate crossing and parking based on computer vision and point cloud renderings, and the use of graph-based Simultaneous Localization and Mapping (SLAM) for mapping and navigation in challenging areas of the SPS, such as the six access points and the beam-dump area. Machine learning algorithms such as OCR are used to understand the robot position in the SPS based on plate references on the machine. In the LHC, machine learning and computer vision techniques are combined to detect Beam Loss Monitor (BLM) sensors around the beamline. The pose estimation of these sensors is then used as target locations for a robotic arm integrated in the Train Inspection Monorail (TIM) robot to calibrate each sensor using a radioactive source. Point cloud imagery is also used to buildup environmental awareness of the tunnel, facilitating obstacle avoidance, robust operation and digital twin possibilities for future interventions.

Speaker: Eloise Matheson (CERN)

20250408 - ICFA ML Workshop.pdf

20250408 - ICFA ML Workshop.pptx

12:15 Application of boosted decision trees in beam anomaly detection with high throughput data streaming system - 10'+5'

The National Synchrotron Radiation Centre SOLARIS is a third generation light source. SOLARIS, as a big science facility with seven fully operational beamlines, is obligated to provide the best possible conditions for conducting research. One of the ways to create favorable environment is delivering precise tools for teams working across many different fields in SOLARIS. The general problem that still remains alive and it is not fully covered is the detection of beam anomalies. The challenge we faced was to find common points among all of synchrotron’s fields and combine them into comprehensive, reliable system which aims to provide continuous operation and improve quality of beam shared for synchrotron users. Solution we would like to present is a system which gathers diagnostics signals (e.g. from BPMs, magnets system, vibration detecting system and other sensors) and ingest them in real-time into machine learning (ML) unit. The current system consists of two parts. The first one is high capable data streaming layer which handles high frequency signal (fast acquisition). The second one is ML unit which uses boosted decision trees (BDT) to perform real-time classification whether beam parameters are good or not and predict future beam dump at the early stage. At this moment the main data source for ML model are signals from Beam Position Monitors. Moreover, algorithms such as Fast Fourier Transform (FFT), Probability Density Function (PDF) or Principal Component Analysis (PCA) are working simultaneously in background and gives diagnostic team factors to make a proper evaluation and take actions before automatic beam dump mechanism will be activated. Additionally, PDF and PCA algorithms give ability to detect exact location of issues.

Speaker: Mr Maciej Mleczko (National Synchrotron Radiation Centre)

data_streaming_system_ICFA_2025.pdf

12:30 An Integrated Research Infrastructure framework for digital twins of laser-plasma acceleration experiments- 10'+5'

Laser-plasma acceleration is a promising acceleration technology for a number of applications due to the large accelerating gradient and unique beam properties that it produces. This technology is in active development, and experimental campaigns typically dedicate significant time to exploring the parameter space in real time, adjusting laser properties, target configuration, and other factors to find the optimal setup. Therefore, having AI/ML-driven digital twins capable of providing real-time optimization guidance, by learning from both experimental and simulation data, would be highly beneficial.

We present progress towards an Integrated Research Infrastructure framework that can run multi-GPU simulations during experimental campaigns, and combine the results with ongoing measurements to provide real-time guidance. The framework is primarily intended for use at LBNL’s BELLA Center and leverages several open-source tools, including WarpX and LASY for simulations, lume-model for the AI/ML digital twin, NERSC’s superfacility API for submitting multi-GPU simulations in real time and the Prefect platform for overall workflow orchestration. Additionally, we will discuss challenges in building ML models that must learn from distinct and potentially mismatched data sources.

Speaker: Dr Remi Lehe (LBNL)

2025_04_07_CERN_conference.pdf

12:45 → 14:00 Lunch

14:00 → 15:00 MLOps, Infrastructure and Scalability

Convener: Tia Miceli

14:00 Lume Deployment - standardising model deployment- 15'+5'

Ensuring efficient use of resources and longevity of machine learning projects requires careful consideration of the full machine learning lifecycle especially when models are deployed to interact with live control systems or end users. We present Lume Deployment a framework of standardised modules built for rapid development and deployment of machine learning models and their integration to the control system. The framework is built around a modular approach which separates the system, the data and the model from each other via well-defined interfaces, which also allows easy expandability of the framework. We showcase several case studies and use-cases that demonstrate the framework’s flexibility in various real-world scenarios. Additionally, we outline current developments and possible future developments focusing on the framework’s stability and automation model evaluation tasks.

Speaker: Mateusz Leputa

Poly-Lithic.pptx

14:20 Software Infrastructure Plans for End-to-End and Bottom-Up AI/ML Capabilities at the Electron-Ion Collider (EIC)- 15'+5'

Plans for the Electron-Ion Collider (EIC), to be built at Brookhaven National Laboratory, include end-to-end and bottom-up capabilities in artificial intelligence (AI) and machine learning (ML). Enabling these capabilities, especially for EIC Operations, will require the large-scale integration of software platforms and tools for the reliable and efficient management of AI/ML-related data, workflows, and processes. We present on our software infrastructure plans.

Speaker: Linh Nguyen (Brookhaven National Laboratory)

Nguyen_EIC_MaLAPA_2025_VersionForUpload.pdf

Nguyen_EIC_MaLAPA_2025_VersionForUpload.pptx

14:40 Machine Learning for the SNS Accelerator and Target- 15'+5'

We apply Machine Learning techniques at the Spallation Neutron Source (SNS) to improve operations, specifically to deter and prevent errant beam pulses, to speed up minimization of halo beam losses, and to alert operators to anomalies in the target cooling system. We give an overview of the work done and discuss the infrastructure implemented and under development to support the data acquisition, pre-processing, training, and continuous learning.

Speaker: Willem Blokland (ORNL)

documents 2025-ICFA-Blokland.pdf (Protected) 2025-ICFA-Blokland.pptx (Protected)

15:00 → 15:30 Coffee break

15:30 → 16:00 Community session: MLOps topic

Conveners: Kishan Rajput (Jefferson Lab), Thorsten Hellert

20250410_ICFA_AIML_Roadmap_Discussion.pdf

20250410_ICFA_AIML_Roadmap_Discussion.pptx

16:00 → 18:00 Poster session: Posters

19:00 → 21:00 Dinner

Thursday 10 April

08:30 → 10:30 Anomaly Detection and Diagnostics

Convener: Annika Eichler (DESY)

08:30 A path to efficient machine learning-based beam diagnostics: complete six-dimensional generative phase space reconstruction without RF deflecting cavity- 15'+5'

Generative phase space reconstruction method based on neural networks and differentiable simulations has become a novel beam diagnostic technique to obtain the beam phase space information. Recent studies show that four-dimensional phase space can be successfully obtained by using only YAG images with different quadrupole magnet strength, allowing us to understand both uncoupled and coupled phase spaces. Furthermore, it has been experimentally demonstrated that the complete six-dimensional phase space can be reconstructed by additionally utilizing a spectrometer dipole magnet and RF transverse deflecting cavity. In addition to the previous research activities, we are currently investigating the complete six-dimensional phase space reconstruction method that does not require the RF transverse deflecting cavity. We demonstrate in simulation that our proposed method can also provide complete six-dimensional phase spaces including all the transverse-longitudinal couplings, which successfully represent the ground truth distributions. In this study, we present how to perform the reconstruction without such an advanced diagnostic instrument. In addition, we show the reconstruction results with synthetic examples and actual experimental data obtained at the Pohang Accelerator Laboratory X-ray Free Electron Laser (PAL-XFEL) facility.

Speaker: Seongyeol Kim

6D_GPSR_PALXFEL_Seongyeol_Kim_Final_Presentation.pdf

08:50 Measurement of CSR effects using generative phase space reconstruction- 15'+5'

Coherent synchrotron radiation (CSR) is a limiting effect in linear accelerators with dispersive elements due to its contribution to projected transverse emittance growth. This effect becomes a limitation for highly compressed beams. Even though CSR-induced projected emittance growth has been widely studied, conventional measurement techniques are not detailed enough to resolve the multi-dimensional structure of the beam, namely the different translations and rotations of transverse phase space slices throughout the longitudinal coordinate. In this work, we use a state-of-the-art method to reconstruct the phase space of a beam affected by CSR at the Argonne Wakefield Accelerator Facility. This detailed, efficient and multi-dimensional phase space reconstruction method enables better understanding of the CSR effects in a double dogleg where shielding is limited.

Speaker: Juan Pablo Gonzalez-Aguilera (University of Chicago)

MaLAPA 2025.pdf

MaLAPA 2025.pptx

09:10 Semi-supervised detection of optics errors in beamlines- 15'+5'

Optics tuning in transfer lines and LINACs can be challenging due to the fact that multiple combinations of machine settings can lead to the same diagnostic output. Moreover, the lack of a periodic solution can limit the ability to infer optics in the same way as rings from BPM signals. Model based approaches are often used to assist with the optics tuning in combination with optimization or parameter estimation. Here we have developed a novel approach using machine learning inverse models trained on a known configuration to detect variations in quadrupole settings without explicitly including them in the model. This paper shows a comparison of neural network models and linear models on both a simulation-based study and experimental studies conducted at the AGS to RHIC transfer line at Brookhaven National Lab.

Speaker: Jonathan Edelen

MaLAPA_JEdelen.pdf

MaLAPA_JEdelen.pptx

09:30 Anomaly Forecasting and Adaptive Learning in Fast Kicker Magnet Systems- 15'+5'

Kicker magnets are essential for particle beam injection and extraction within CERN’s accelerator complex, where high reliability is crucial to maintaining the availability needed for numerous scientific experiments. This study proposes a machine learning approach for forecasting anomalies in these systems, aiming to proactively identify and isolate potential faults before failure occurs. To keep the anomaly detection model accurate over time, continual learning techniques are employed, allowing the model to adapt to evolving system dynamics without frequent retraining. This combination enhances the efficiency and stability of accelerator operations by ensuring the model remains up-to-date in the face of non-static data.

Speaker: Malik Marco Algelly (Universite de Geneve (CH))

2025_ICFA_Algelly.pdf

2025_ICFA_Algelly.pptx

09:50 Enhancing Quench Detection in SRF Cavities at the European XFEL: Machine Learning Approaches and Practical Challenges- 15'+5'

At the European XFEL, detecting anomalies in superconducting cavities is essential for reliable accelerator performance. We began with a model-based fault detection approach focused on residual analysis to identify anomalies. To improve fault discrimination, particularly for quench events, we augmented this system with machine learning (ML) models. Key challenges included the scarcity of labeled data, which we addressed by integrating expert feedback through an optimized labeling process, and the transition to real-time operation, requiring computational and integration adjustments. For the online application, we deployed two servers in the tunnel at one of the 25 stations to detect failures in real-time with a software-based solution. In parallel we pushed the development of an FPGA-based solution, that will allow to counteract on real-time in the future. The resulting detection system delivers reports across various timescales, supporting both immediate responses and long-term maintenance. It will provide new insights to the online data, which was never explored in the past.

Speakers: Lynda Boukela (DESY), Burak Dursun (DESY)

Malapa2025 - Boukela.pdf

Malapa2025 - Boukela.pptx

10:10 Improving Coincident Learning for Beam-based RF Station Fault Identification Using Phase Information at the SLAC Linac Coherent Light Source- 15'+5'

The vast amount of data generated by accelerators makes manual monitoring impractical due to its labor-intensive nature. Existing machine learning solutions often rely on labeled data, manual inspection, and hyperparameter tuning, which limits their scalability. To address these challenges, we leverage coincidence learning—an unsupervised technique designed for multi-modal tasks—to automatically detect anomalies by identifying coincident patterns of behavior across two distinct segments of the feature space. Specifically, we focus on anomaly detection for radio-frequency (RF) stations at the SLAC Linac Coherent Light Source (LCLS). By analyzing shot-to-shot data from the beam position monitoring system alongside data from RF stations, we can identify the source of changes in the accelerator's status. Previous studies on RF stations produced reasonable results using time-asynchronous amplitude data, but ignored the richer information from time-synchronous phase data due to its complexity. We find that using neural networks to analyze the phase data enables the detection of anomalies that amplitude-based detection missed. Additionally, the time-synchronous phase data provides critical insights, allowing us to distinguish whether an RF station change occurs simultaneously with changes in the accelerator status or in response to them. Additionally, the rich information contained in the phase data facilitates clustering of anomalies into distinct categories, each with unique signatures. This categorization brings us closer to identifying the root causes of issues within the RF stations.

Speaker: Jason Liang (Stanford University)

COAD_Beam_Revised.pdf

COAD_Beam_Revised.pdf

10:30 → 11:00 Coffee Break

11:00 → 12:20 Surrogate Modelling and Digital Twins

Convener: Francesco Maria Velotti (CERN)

11:00 Graph Learning for Explainable Operation of Particle Accelerators- 15'+5'

We describe research in deep learning on graph representations of the injector beamline at the Continuous Electron Beam Accelerator Facility (CEBAF) to develop a tool for operations. We leverage operational archived data – both unlabeled and labeled configurations – to train a graph neural network (GNN) via our methods of self-supervised training and supervised fine tuning. We demonstrate the ability of the GNN to distill high-dimensional beamline configurations into low-dimensional embeddings and use them to create an intuitive visualization for operators. By mapping out regions of latent space characterized by good and bad setups, we describe how this could provide operators with more informative, real-time feedback during beam tuning compared to the standard practice of interpreting a set of sparse, distributed diagnostic readings. We further describe the results of a framework that provides users with explanations for why a configuration changes location in the latent space.

Speaker: Chris Tennant

ML_WORKSHOP_2025_TENNANT_FINAL.pptx

11:20 Compensating hysteresis for the CERN SPS main magnets with transformers- 15'+5'

This work presents a machine learning-based approach for compensating magnetic hysteresis in the main dipole and quadrupole magnets of the multi-cycling CERN SPS, utilizing time series neural architectures like the Temporal Fusion Transformers trained on magnetic field measurements. The predicted magnetic fields enable feed-forward, cycle-by-cycle, compensation through the CERN accelerator control system, thereby improving beam reproducibility and enhancing operational flexibility. Correcting deterministically for hysteresis effects removes the necessity of energy-, and time-consuming magnet-precycling while guaranteeing reproducible beam parameters in the SPS. Achieving a magnetic field prediction accuracy in the range of 5e-5 T, this approach demonstrates significant potential for beam optimization, energy savings and improved sustainability, with potential for implementation for any synchrotron magnet circuit. In addition to addressing the technical implementation and benefits, the challenges associated with real-time autoregressive application, model accuracy, and performance evaluation for cycle-to-cycle operation without feedback will be discussed.

Speaker: Anton Lu (Technische Universität Wien (AT))

lua_ICFA_20250410_hystcomp.pdf

lua_ICFA_20250410_hystcomp.pptx

11:40 Improving fast beam transport simulations using transfer learning- 15'+5'

Fast simulations of intense relativistic electron beams can be sufficiently accurate to allow for tuning of an accelerator’s magnetic transport field, but are incapable of capturing all relevant beam physics due to limitations in the model. Because methods that do capture these effects are significantly more computationally-expensive, e.g. particle in-cell simulations, they are fundamentally less useful for optimization problems. Here, transfer learning with high-fidelity, full-physics models is applied to the output of a machine learning model trained on a dataset generated by a fast particle beam simulation to bring the simulation results more in line with experimental data.

This work was done by Mission Support and Test Services, LLC, under Contract No. DE-NA0003624 with the U.S. Department of Energy, and the National Nuclear Security Administration. DOE/NV/03624--2062.

Speaker: Paul Stanik III (University of Nevada--Las Vegas)

stanikp_5th_icfa_2025_16x9_final.pdf

stanikp_5th_icfa_2025_16x9_final.pptx

12:00 Optimizing High-Energy Beam Transport with AI: Advances in IFMIF-DONES Design- 15'+5'

A major challenge in constructing future nuclear fusion power plants is understanding how reactor materials are damaged by the neutron flux generated during the fusion process. In order to address this challenge, the IFMIF-DONES neutron source is being built for material irradiation, generating the necessary neutron flux through a stripping reaction between accelerated deuterons and a lithium circuit. In this work, within the DONES-FLUX project, Fourier Neural Operators are employed as Deep Learning Surrogate Models for optimizing the design of the High-Energy Beam Transport Line of the IFMIF-DONES accelerator. The trained models, which predict the deuteron beam statistical functions and beam profile distributions, are roughly 3 orders of magnitude faster than traditional simulations while keeping mean absolute percentage errors below 5%. This significant reduction in inference time, along with the models' differentiability, enables the use of optimization algorithms like online Reinforcement Learning, Bayesian Optimization, and Gradient Descent. Additionally, this last method was implemented and tested for finding optimal quadrupole values for different beam configurations, where solutions are reached within minutes. These positive results highlight the synergy between different Deep Learning architectures and offer a promising collaboration between the field of Artificial Intelligence and accelerator facilities.

Speakers: Galo Gallardo Romero (HI Iberia), Guillermo Rodríguez Llorente (Artificial Intelligence Engineer at HI Iberia, Department of Mathematics at UC3M, Gregorio Millán Barbany Institute)

Optimizing High-Energy Beam Transport with AI Advances in IFMIF DONES Design.pdf

Optimizing High-Energy Beam Transport with AI Advances in IFMIF DONES Design.pptx

12:20 → 14:00 Lunch

14:00 → 15:00 Community session: Framework for Digital Twins

Conveners: Georg Hoffstaetter, Remi Lehe

20250410_ICFA_AIML_Roadmap_Discussion.pdf

20250410_ICFA_AIML_Roadmap_Discussion.pptx

15:00 → 15:30 Coffee break

15:30 → 16:50 Surrogate Modelling and Digital Twins

Convener: Mr Jiao Yi (IHEP)

15:30 End-to-end differentiable digital twin for the IOTA/FAST facility- 15'+5'

As the design complexity of modern accelerators grows, there is more interest in using controllable-fidelity simulations that have fast execution time or yield additional insights as compared to standard codes. 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 an end-to-end 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 for linac and ring simulation, the development of generic interfaces between surrogate and physics-based sections, and presenting the control interface as either a deterministic event loop or a fully asynchronous EPICS soft IOC. We will also discuss challenges in model calibration and uncertainty quantification, as well as future plans to extend modelling to larger machines like PIP-II and Booster.

Speaker: Nikita Kuklev (Fermilab)

icfaml25_nk_twin_talk.pdf

15:50 An Online Virtual Model for the ATLAS Ion Linac at Argonne- 15'+5'

A multifaceted virtual accelerator model that seamlessly integrates with the online experimental system would highly benefit the operators to test and evaluate beam tuning scenarios and apply them online. As part of this effort, the beam dynamics code TRACK is wrapped with control system architecture and the graphic user interface BADGER developed by SLAC. Customizability and task visualization are prioritized based on end user feedback. More important than the user experience is how well the simulation model agrees with the experimental measurements. A recent effort to account for beam steering and misalignment effects has reduced the error between simulation and measurements from ~ 40% to ~ 10%. More work to pare down this error and close the gap even further is currently underway. The concurrent execution of simulation with experiment for troubleshooting and parallel simulation-based optimizations to quickly determine ideal element settings that map to desired beam parameters are some of the features to be highlighted. Previously published AI/ML capabilities are incorporated to this extent. Surrogate models based on collected data to reproduce a given set of operating conditions (beam energy, charge state etc.) and use of an LLM to monitor simulation progress and make informed decisions at a higher level will be included as add-on features.

Speaker: Adwaith Ravichandran (Argonne National Laboratory)

ICFA_Adwaith_Ravichandran_Virtual_Model.pptx

ICFA-ML-Abstract_Adwaith_Ravichandran.docx

16:10 ML-Based Multi-Fidelity Model Calibration Toward Precision Control of Electron Beams- 15'+5'

Precision control of electron beams is one of the main charges of beam physics, as producing high-brightness beams is critical to numerous accelerator deliverables, including high-quality x-rays from XFELs and high-quality ultrafast probes for UED/UEM. Critical to this effort is a set of accurate system models that can inform control policies. To be useful, these models must accurately reflect the behavior of the accelerator. In this work, a systematic, ML-based approach toward this model calibration problem is outlined. We use ML-based, time-efficient approaches, such as multi-fidelity Bayesian optimization, to balance the flow of information from high- and low-fidelity models. Additionally, the application of this work to online digital twins toward higher brightness beams will be discussed.

Speaker: Eric Cropp

2025-04-10_ICFA_Final.pdf

16:30 Developing Radiation-Tolerant Transverse Beam Imaging Using Synthetic Data and Multimode Fiber- 15'+5'

Beam imaging presents significant challenges due to the necessity of positioning imaging devices near the beam pipe, an area subjected to high levels of radiation that can damage cameras and their peripheral electronics, reducing their lifespan and reliability. With the global discontinuation of radiation-hardened tube cameras previously used for this purpose, a robust and durable replacement imaging solution is needed. Multimode optical fibers have emerged as viable alternatives, capable of relaying the image signal to a standard CMOS camera location in a radiation-safe environment. A challenge within this approach is mode coupling and scattering within the fiber, which increases the difficulty in accurately reconstructing beam information.

This contribution showcases a method of reconstructing transverse beam distribution parameters from a distorted fiber output. This is achieved with an experimental setup that makes use of a synthetic input dataset generated from multiple high-variance 2D Gaussian fields, multimode optical fibers to propagate and distort these images, and a 2D convolutional autoencoder to reconstruct the inputs. This setup is used as a training dataset, with the input dataset chosen to support generalizability. Our machine learning model is tested on a real dataset of transverse beam distributions collected at CERN’s CLEAR facility. We achieve an average RMSE of 2.44% over four key transverse beam parameters after reconstruction on the testset. Our model further demonstrates mitigated bias in beam parameter estimation and strong generalization capability, reconstructing well radically different parameter distributions.

Speaker: Qiyuan Xu

ICFA_2025_QiyuanXu.pdf

ICFA-Poster-QiyuanXu.pdf

16:50 → 17:50 Community session: AI roadmap topic

Conveners: Auralee Edelen, Verena Kain (CERN)

20250410_ICFA_AIML_Roadmap_Discussion.pdf

20250410_ICFA_AIML_Roadmap_Discussion.pptx

18:15 → 19:30 Bus transportation to Geneva city center for the Gala dinner

19:30 → 22:30 Dinner at Beau Rivage Hotel

22:30 → 23:00 Bus transportation to Cern B39

Friday 11 April

08:30 → 09:30 Large Language Model Chatbots for Enhanced Documentation Access

This tutorial applies Retrieval Augmented Generation (RAG) as a method to improve documentation retrieval in accelerator physics. Participants will learn how combining information retrieval with generative AI models can provide precise, context-aware answers from vast technical resources. The session includes a hands-on demonstration of implementing RAG in combination with Large Language Models (LLMs) and retrieving information from your documentation.

Speaker: Dr Florian Rehm (CERN)

01_RAG_Intro.html

01_RAG_Intro.ipynb

01_RAG_Intro.pdf

02_Agent_Intro.html

02_Agent_Intro.ipynb

02_Agent_Intro.pdf

ICFA_Tutorial_LLMs_FlorianRehm_v1_3.pdf

ICFA_Tutorial_LLMs_FlorianRehm_v1_3.pdf

ICFA_Tutorial_LLMs_FlorianRehm_v1_3.pptx

09:30 → 09:50 LLMs and AI Assistants: sulz

09:30 Multi-modal LLMs at EuXFEL: Knowledge Assistants, Data Exploration, and Policy- 15'+5'

Multimodal Large Language models extend LLMs’ capabilities to input beyond text, often images. At the European XFEL, these models are used as Retrieval-Augmented Generative (RAG) Knowledge assistants in technical and administrative domains. We present a selection of current applications and prototypes: chatbot assistants for data service support, business travel aid, vision-based document exploration systems, coding assistants, and an application rating control system panels for accessibility and design. The aforementioned examples were studied as part of an effort to create a policy for the usage of LLMs by facility staff. We highlight how this policy is influenced by the aforementioned examples and associated legal and data privacy considerations.

Speaker: Florian Sohn (European XFEL GmbH)

LLM_ICFA.pdf

LLM_ICFA.pptx

09:50 → 10:20 Group photo and Coffee break

10:20 → 11:00 LLMs and AI Assistants

10:20 The Journey of Developing Specialized Text Embedding Models- 15'+5'

The specialized terminology and complex concepts inherent in physics present significant challenges for Natural Language Processing (NLP), particularly when relying on general-purpose models. In this talk, I will discuss the development of physics-specific text embedding models designed to overcome these obstacles, beginning with PhysBERT—the first model pre-trained exclusively on a curated corpus of 1.2 million arXiv physics papers. Building upon this foundation, we turn our attention to accelerator physics, a subfield with even more intricate language and concepts. To effectively capture the nuances of this domain, we developed AccPhysBERT, a sentence embedding model fine-tuned specifically for accelerator physics literature. A key aspect of this development involved leveraging Large Language Models (LLMs) extensively to generate annotated training data, enabling AccPhysBERT to facilitate advanced NLP applications such as semantic paper-reviewer matching and integration into Retrieval-Augmented Generation systems.

Speaker: Thorsten Hellert

250411_MALAPA.pdf

10:40 Towards Agentic AI on Particle Accelerators- 15'+5

As particle accelerators grow in complexity, traditional control methods face increasing challenges in achieving optimal performance. This paper envisions a paradigm shift: a decentralized multi-agent framework for accelerator control, powered by Large Language Models (LLMs) and distributed among autonomous agents. We present a proposition of a self-improving decentralized system where intelligent agents handle high-level tasks and communication and each agent is specialized control individual accelerator components.

This approach raises some questions: What are the future applications of AI in particle accelerators? How can we implement an autonomous complex system such as a particle accelerator where agents gradually improve through experience and human feedback? What are the implications of integrating a human-in-the-loop component for labeling operational data and providing expert guidance? We show two examples, where we demonstrate viability of such architecture.

Speaker: Raimund Kammering

250411_MALAPA_Agents.pdf

250411_MALAPA_Agents.pdf

250411_MALAPA_Agents.pptx

250411_MALAPA_Agents.pptx

11:00 → 12:00 summary, preview, final words

Speakers: Auralee Edelen, Daniel Ratner (SLAC), Georg Hoffstaetter, Hirokazu Maesaka, Jan Kaiser, Kishan Rajput (Jefferson Lab), Remi Lehe, Ryan Roussel, Thorsten Hellert, Verena Kain (CERN)

Final.pdf

lua_ICFA_visits.pdf

lua_ICFA_visits.pptx

MaLAPA26AnnounceMaesaka.pdf

malapa-discord.pdf

Summary of MaLAPA 2025 Discussion Sessions.pdf

12:00 → 13:30 Lunch

13:30 → 16:30 CERN visits