Particle accelerator optimization, so finding optimal operation points is a challenging task. Expensive physics models, like OPAL [1], are usually applied for that task. To overcome the computational burden of such models, two kinds of data-driven models, [2], are presented in this talk. The first one, is based on a densely connected neural network. It models the machine at any longitudinal position, from machine settings to final beam parameters. The second model is based on Invertible Neural Networks, [3], and it simulates the forward and the inverse pass, so from machine settings to beam parameters and vice versa. Further, the application of both models for multi-objective particle accelerator optimization and the according speed-up is demonstrated.

[1] Adelmann, A.; Calvo, P.; Frey, M.; Gsell, A.; Locans, U.; Metzger-Kraus, C.; Neveu, N.; Rogers, C.; Russell, S.; Sheehy, S.;Snuverink, J.; Winklehner, D. OPAL a Versatile Tool for Charged Particle Accelerator Simulations.arXiv e-prints2019, arXiv:1905.06654, [physics.acc-ph].

[2] Bellotti, R.; Boiger, R.; Adelmann, A. Fast, efficient and flexible particle accelerator optimisation using densely connected and invertible neural networks. arXiv e-prints2021, arXiv:2107.00060, [physics.acc-ph].

[3] Ardizzone, L.; Kruse, J.; Wirkert, S.; Rahner, D.; Pellegrini, E.W.; Klessen, R.S.; Maier-Hein, L.; Rother, C.; Köthe, U.Analyzing Inverse Problems with Invertible Neural Networks.arXiv e-prints2018, arXiv:1808.04730v3[LG, ML].

The Low Energy RHIC electron Cooling (LEReC) system is the world's first electron cooler utilizing radio frequency (RF) accelerated electron bunches and a non-magnetized electron beam. It is also the first electron cooler applied directly to colliding hadron beams. The unique approach to cooling makes beam dynamics in LEReC very different from the conventional electron coolers. Numerous LEReC parameters can affect the cooling rate. One of the most critical factors is the alignment of the electron and ion trajectories in the cooling section. In this work, we apply Bayesian optimization to check and if needed to optimize the trajectories' alignment. Experimental results are presented, and it is demonstrated that machine learning (ML) methods can be applied to perform the control tasks effectively in the RHIC controls system.

Bio: Yuan Gao is an assistant scientist of the Control Systems groups in the Collider Accelerator Department (C-AD) at Brookhaven National Laboratory (BNL), which operates the Relativistic Heavy Ion Collider (RHIC). He also organizes a seminar series in the Controls group where people discuss Machine Learning (ML) applications in the system. Currently, his research focus is on the use of ML, game theory techniques in accelerator control systems as well as developing the infrastructure for control system simulations.

Lucy Lin is a third-year Ph.D. student at Cornell University working with Prof. Georg Hoffstaetter. Her main research interests are focused on exploring applications of machine learning (ML) algorithms in accelerator operations. She has been working with groups at Brookhaven National Laboratory (BNL) to develop machine learning algorithms that improve the performances of cooling experiments at the Relativistic Heavy Ion Collider (RHIC), including Low Energy RHIC electron Cooling (LEReC) and Coherent electron Cooling (CeC).

The beam interruptions (interlocks) of particle accelerators, despite being necessary safety measures, lead to abrupt operational changes and a substantial loss of beam time. Novel time series classification approaches are applied to decrease beam time loss in the High-Intensity Proton Accelerator complex by forecasting interlock events. The forecasting is performed through binary classification of single timestamps as well as windows of multivariate time series, with methods ranging from Lasso models to Recurrence Plots followed by Convolutional Neural Network. Our best-performing interlock-to-stable classifier reaches an Area under the ROC Curve value of 0.71±0.01, and it can potentially reduce the beam time loss by 0.5±0.2 s per interlock.

Virtual Diagnostic (VD) is a computational tool built using deep learning that can be used to predict diagnostic outputs. VDs are especially useful in systems where measuring outputs is invasive, limited, costly or runs the risk of modifying the output. In experiments with large ramifications, it is important to quantify the uncertainty of each prediction. Given out-of-distribution inputs (e.g. using the same machine in a different operation mode), it is also necessary to understand how robust the VD model is and how well it generalizes on unfamiliar data. In this work, we use various compositions of neural networks to explore and enhance prediction uncertainty and robustness on data sets gathered from SLAC National Laboratory. We aim to accurately and confidently predict the longitudinal phase space images of electron beams. The ability to make informed decisions under uncertainty and limited computational power is crucial for reliable deployment of scalable deep learning tools on safety-critical systems such as particle accelerators

Machine learning (ML) tools are able to learn relationships between inputs and outputs of large complex systems directly from data. For time-varying systems, the predictive capabilities of ML tools degrade if the systems are no longer accurately represented by the data with which the ML models were trained. For large systems, re-training is only possible if the changes are slow relative to the rate at which large numbers of new input-output training data can be non-invasively recorded. We present several adaptive machine learning (AML) approaches to deep learning for time-varying systems without re-training. We demonstrate our methods with several examples of applications and simulation studies for time-varying particle accelerator beams and components including AML for controlling the longitudinal phase space of the electron beam at the LCLS FEL [1], an AML-based inverse model for mapping output beam measurements to input beam distributions at the HiRES UED [2], and a simulation study of an AML method for non-invasive beam envelop size diagnostics at the LANSCE linear ion accelerator [3].

[1] A. Scheinker, et al. "Demonstration of model-independent control of the longitudinal phase space of electron beams in the linac-coherent light source with femtosecond resolution." Physical review letters 121.4 (2018): 044801.

DOI: https://doi.org/10.1103/PhysRevLett.121.044801

[2] A. Scheinker, et al. "Adaptive Deep Learning for Time-Varying Systems With Hidden Parameters: Predicting Changing Input Beam Distributions of Compact Particle Accelerators." arXiv preprint arXiv:2102.10510 (2021).

DOI: 10.21203/rs.3.rs-373311/v1

[3] A. Scheinker. “Adaptive Machine Learning for Robust Diagnostics and Control of Time-Varying Particle Accelerator Components and Beams.” Information. 2021; 12(4):161.

DOI: https://doi.org/10.3390/info12040161

Machine learning (ML) tools are able to learn relationships between inputs and outputs of large complex systems directly from data. For time-varying systems, the predictive capabilities of ML tools degrade if the systems are no longer accurately represented by the data with which the ML models were trained. For large systems, re-training is only possible if the changes are slow relative to the rate at which large numbers of new input-output training data can be non-invasively recorded. We present several adaptive machine learning (AML) approaches to deep learning for time-varying systems without re-training. We demonstrate our methods with several examples of applications and simulation studies for time-varying particle accelerator beams and components including AML for controlling the longitudinal phase space of the electron beam at the LCLS FEL [1], an AML-based inverse model for mapping output beam measurements to input beam distributions at the HiRES UED [2], and a simulation study of an AML method for non-invasive beam envelop size diagnostics at the LANSCE linear ion accelerator [3].

[1] A. Scheinker, et al. "Demonstration of model-independent control of the longitudinal phase space of electron beams in the linac-coherent light source with femtosecond resolution." Physical review letters 121.4 (2018): 044801.

DOI: https://doi.org/10.1103/PhysRevLett.121.044801

[2] A. Scheinker, et al. "Adaptive Deep Learning for Time-Varying Systems With Hidden Parameters: Predicting Changing Input Beam Distributions of Compact Particle Accelerators." arXiv preprint arXiv:2102.10510 (2021).

DOI: 10.21203/rs.3.rs-373311/v1

[3] A. Scheinker. “Adaptive Machine Learning for Robust Diagnostics and Control of Time-Varying Particle Accelerator Components and Beams.” Information. 2021; 12(4):161.

DOI: https://doi.org/10.3390/info12040161

The dynamic aperture (DA) is an important concept in the study of nonlinear beam dynamics. Several analytical models used to describe the evolution of DA as a function of time, and to extrapolate to realistic time scales that would not be reachable otherwise due to computational limitations, have been successfully developed. Even though these models have been quite successful in the past, the fitting procedure is rather sensitive to several details. Machine Learning (ML) techniques, which have been around for decades and have matured into powerful tools ever since, carry the potential to address some of these challenges. In this paper two applications of ML approaches are presented and discussed in detail. Firstly, ML has been used to efficiently detect outliers in the DA computations. Secondly, ML techniques have been applied to improve the fitting procedures of the DA models, thus improving their predictive power.

The incorporation of Machine Learning tools is a central part of the next years' roadmap at the large-scale user facility BESSY II (operated by the Helmholtz-Zentrum Berlin). In order to improve the performance and increase the machine automation, Reinforcement Learning (RL) agents have been already developed a tested at BESSY II for several use cases such as booster current or injection efficiency optimisation. In this talk we will focus mainly on the the application of model-free RL agents for the mitigation of orbit perturbations in the storage ring. In particular, we are interested in harmonic perturbations produced by the environment - for example due to the main power or some imperfectly isolated magnetic sources. We will cover the methodology, design and simulation phases as well as the challenges faced and the first results obtained during our tests at the machine.

We report on the development of machine learning models for the recognition, identification, and prediction of C100 superconducting radio-frequency (SRF) cavity faults in the Continuous Electron Beam Accelerator Facility (CEBAF) at Jefferson Lab. CEBAF is a continuous-wave recirculating linac utilizing SRF cavities to accelerate electrons up to 12 GeV. The C100 SRF cavities in CEBAF are designed with a digital low-level RF system configured to retain waveform recordings in a cavity failure event. Subject matter experts (SME) are able to analyze the collected time-series recordings and determine the type of fault, and the offending cavity. This information is used to identify failure trends and apply corrective measures to the problematic cavity. However, manual analysis of large-scale RF data to identify the cavity and the fault type is laborious and time-consuming. Consequently, we develop several machine learning and deep learning models to automate the process of cavity and fault classification with near-real time recognition capability. We discuss the performance of these models using an RF waveform dataset built using the past runs of the CEBAF, and present a real-world performance analysis on a model deployed in CEBAF through a recent physics run. Additionally, we discuss research efforts into the potential discovery and categorization of fault types through unsupervised machine learning techniques, and present preliminary work on the feasibility of cavity and fault prediction using RF data collected prior to a failure event.

In state-of-the-art synchrotron light sources the overall source stability is limited by the achievable level of electron beam size stability. This source size stability is presently on the few-percent level, which is still 1-2 orders of magnitude larger than already demonstrated stability of source position/angle and current. Until now, source size stabilization has been achieved through corrections based on a combination of static predetermined physics models and lengthy calibration measurements, periodically repeated to counteract drift in the accelerator and instrumentation. We now demonstrate for the first time [PRL 123 194801 (2019)], how application of machine learning allows for a physics- and model-independent stabilization of source size relying only on previously existing instrumentation in ALS. Such feed-forward correction based on neural networks that can be continuously online-retrained achieves source size stability as low as 0.2 microns rms (0.4%) which results in overall source stability approaching the sub-percent noise floor of the most sensitive experiments.

This talk discusses the implementation of a suite of virtual diagnostics for electron beam prediction and control with an emphasis on recent progress and planned work at the FACET-II facility currently under commissioning at SLAC National Accelerator Laboratory. The diagnostics will be used for prediction of the longitudinal phase space along the linac, spectral reconstruction of the bunch profile and non-destructive inference of transverse beam quality (emittance) using edge radiation at the bunch compressor locations. These measurements will be folded in to adaptive feedbacks and ML-based reinforcement learning controls to improve the stability and optimize the performance of the machine for different experimental configurations. I will describe each of these diagnostics with achieved/expected measurement results based on simulation and experimental data and discuss progress towards implementation in regular operations.

Recently, the application of ML has grown in accelerator physics, in particular in the domain of diagnostics and control. One of the first applications of ML at the LHC is focused on optics measurements and corrections. Unsupervised Learning has been applied to automatic detection of beam position monitors faults to improve optics analysis, demonstrating successful results in operation. A novel ML-based approach for the estimation of magnet errors is developed, using supervised regression models trained on a large set of LHC optics simulations. Also, auto encoder neural networks have found their application in denoising of measurements data and reconstruction of missing data points. The results and future plans for these studies will be discussed following a brief introduction to relevant ML concepts

The application of machine learning (ML) techniques for anomaly detection in particle accelerators has gained popularity in recent years. These efforts have ranged from the analysis of quenches in RF cavities [1, 2] and superconducting magnets [3] to anomalous beam position monitors [4], and even losses in rings [5]. Using ML for anomaly detection can be challenging owing to the inherent imbalance in the amount of data collected during normal operations as compared to during faults. Additionally, the data are not always labeled and therefore supervised learning is not possible. Autoencoders, neural networks that form a compressed representation and reconstruction of the input data, are a useful tool for such situations. Here we explore the use of autoencoder reconstruction analysis for anomaly detection in two problem domains: root cause analysis for changes in the output current in the Fermilab LINAC and prediction of magnet faults in the APS storage ring.

[1] A. S. Nawaz, S. Pfeiffer, G. Lichtenberg, and H. Schlarb, “Self-organzied critical control for the european xfel using black box parameter identification for the quench detection system,” in 2016 3rd Conference on Control and Fault-Tolerant Systems (SysTol), Sep. 2016, pp. 196–201.

[2] A. Nawaz, S. Pfeiffer, G. Lichtenberg, and P. Rostalski, “Anomaly detection for the european xfel using a nonlinear parity space method,” IFAC-PapersOnLine, vol. 51, no. 24, pp. 1379 – 1386, 2018, 10th IFAC Symposium on Fault Detection, Supervision and Safety for Technical Processes SAFEPROCESS 2018.

[3] M. Wielgosz, A. Skoczea, and M. Mertik, “Using lstm recurrent neural networks for monitoring the lhc superconducting magnets,” Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, vol. 867, pp. 40 –50, 2017.

[4] Elena Fol, “Evaluation of Machine Learning Methods for LHC Optics Measurements and Corrections Software” CERN-THESIS-2017-336, Aug 2017

[5] G. Valentino, R. Bruce, S. Redaelli, R. Rossi, P. Theodoropoulos, and S. Jaster-Merz, “Anomaly detection for beam loss maps in the large hadron collider,” Journal of Physics: Conference Series, vol. 874, p. 012002, jul 2017.

The Large Hadron Collider at CERN makes use of a complex collimation system made up of around 100 collimators to protect its sensitive equipment from potentially dangerous beam halo particles. The collimators are set up around the beam following a multi-stage transverse setting hierarchy. The insertion position of each collimator is established using beam-based alignment techniques, to determine the local beam position and beam size at each collimator. Such alignment techniques involve moving the collimator jaws towards the beam, whilst observing the beam losses in nearby Beam Loss Monitoring (BLM) devices.

In previous years, collimator alignments were performed semi-automatically, with collimation experts present to oversee and control the alignment. During Run II, a new fully-automatic alignment tool was developed, which replaced the user tasks with dedicated algorithms.

This seminar will focus on the user task that was replaced using machine learning models. The suitability of using machine learning during LHC operation was confirmed in 2018, as this new fully-automatic tool was used for all collimator alignments throughout the year. This research has successfully decreased the time required to align collimators by a factor of three and has paved the way for introducing other alignment techniques.

The online operation of particle accelerators represents a complex, black box optimization problem with many inputs and limited, noisy, and temporally-expensive measured outputs. On the other hand, through the use of physics simulations, we often have some prior knowledge of how the accelerator behaves as input parameters are changed. Bayesian surrogate models paired with the appropriate acquisition function can rapidly find the global optimum point in high dimensional parameter spaces with a minimal number of physical measurements. However, accelerator tuning often attempts to simultaneously optimize multiple aspects of beam transport at once. Unfortunately, optimizing certain objectives comes at the expense of others, resulting in a set of pareto-optimal points, known as the pareto front, that optimizes the trade off between objective goals. In this case, acquisition functions that aim to directly increase the pareto front hypervolume can be used with a Bayesian surrogate model to predict the likelihood that a given point in parameter space improves the pareto-optimal set hypervolume. We plan to combine this method, along with a number of other methods that account for hard and soft constraints, to rapidly find the Argonne Wakefield Accelerator drive beam pareto front during operation. In this case, we start to look at how this method can be incorporated into a flexible control system architecture that directly incorporates prior measurements and multi-fidelity simulation data.

Beam instrumentation such as beam loss monitors allow physicists to identify localised or global issues with the operation of the machine. In particular, they are a key diagnostic for movable devices such as collimators. On the other hand, in simulations the study of the dynamic aperture is useful to quantify the impact of the field quality of the magnetic elements of the accelerator's lattice on the accelerator performance. This talk will review the efforts in applying ML in the beam physics group at CERN for the LHC, and focus in particular on the use of anomaly detection methods to detect issues in the beam collimation hierarchy and to detect outliers in dynamic aperture simulations for a given seed, angle and machine configuration. In addition, the use of Gaussian Processes to improve the analytic fitting performance for the evolution of dynamic aperture will be presented

Numerical optimisation algorithms have become part of the toolbox in the control room to increase and stabilise the performance of particle accelerators. These algorithms have many advantages, are available out of the box and can be adapted to any optimisation problem in operation. The next boost in efficiency is expected from Reinforcement Learning algorithms that learn the optimal policy for a certain control problem and in this way can do without the necessary and time-consuming exploration phase of numerical optimisers once trained. Continuous model-free as well as model-based reinforcement learning was successfully tested at various facilities at CERN with up to 16 degrees of freedom. The approach and algorithms that were used will be discussed with the remaining challenges. Finally results from the AWAKE electron line and LINAC4 along with the necessary next steps will be presented.

Online optimization is necessary in particle accelerators because, although physics models exist, there are often significant differences between the simulation and the machine. Machine-learning model-based optimization methods may be beneficial to improve the quality of solution, the speed of convergence and robustness to noise. In this talk, I'll describe the physics-informed Gaussian-Process. We use the physics model's basis function to construct the full model prior including correlations, rather than empirical model selection using historical data. This method is faster to compute, could be easily adapted to other systems, and would be applicable for automatic tuning of new machines/configurations without historical data.

Parameter tuning is a notoriously time-consuming task in accelerator facilities, especially for FELs. What makes optimization at an FEL difficult is typically the noisy environment and the large number of tuning parameters (>20). At SwissFEL, we are using adapted Bayesian optimization by breaking down the global problem into sequential subproblems that can be solved efficiently with safe Bayesian optimization. This allows us to trade off local and global convergence and to adapt to additional structure in the objective function.

At the high intensity proton facility (HIPA) a lot of manual tuning is required to keep the losses at bay which are one of the main limitations in performance and safety. Here the safe Bayesian optimization framework is also used to decrease the losses without running into beam interlocks.

The talk will give an overview of the optimisation framework and it’s use at PSI that has been developed in collaboration with the “Adaptive Systems Group” at ETHZ.