Choose timezone
Your profile timezone:
Powered by Indico
v3.3.2-pre
The fields of large scale data analytics and machine learning have made impressive progress in recent years. Many applications have been successful in applying techniques in these fields for problems in areas such as health, language processing, search engines, etc Many tools have been developed to facilitate the application of these techniques (e.g. libraries like Scikit-learn, TensorFlow, Keras, PyTorch, etc or frameworks like Apache Spark, Caffe, etc)
Although some examples exist of applications in accelerators and experimental physics installations, there is a feeling that we could benefit more from these methods and tools. The workshop is intended to give a tutorial introduction to machine learning and to bring up discussions on experiences and possible applications of advanced data science and machine learning techniques to experimental physics facilities.
The workshop will last one full day. In the morning, introductory tutorials to machine learning will be presented. In the afternoon, speakers are welcome to share their experience with presentations/demonstrations of solutions that worked or didn’t worked well. A final discussion will take place on possible next steps.
Correlated topics: data analytics, statistical analysis, data mining, deep learning, neural networks, expert systems, automatic optimization, robotics, etc.
Invited lecturers: Alfredo Canziani (New York University), Gianluca Valentino (University of Malta)
ICALEPCS 2019 Official Page: https://icalepcs2019.bnl.gov
ICALEPCS 2019 Workshops: https://icalepcs2019.bnl.gov/workshops.html#12
Data science is suite of tools that can be applied to systems. We briefly review the generalities of data science, give several examples of uses in the physical sciences and in the engineering of systems, and discuss recent applications to complex systems such as in accelerator science and engineering. Specifically, we will give several examples where data science including machine learning can be used for the control of scientific systems. Further, we will point the attendee to the many resources in our global community to help better guide them to their own solutions. Finally, we will discuss plans and ideas for future enablers of data science that can be applied to systems such as particle accelerators and laser facilities and well as their research output
Particle accelerators are large, complex and time-varying machines with limited diagnostics and time-varying and uncertain charged particle beams, making it difficult to perform automatic and model-based tuning. Free electron lasers (FEL) and plasma wakefield accelerators (PWFA) are creating more and more complicated electron bunch configurations, including multi-color modes for FELs such as LCLS and LCLS-II and custom tailored bunch current profiles for PWAs such as FACET-II. These accelerators are also producing shorter and higher intensity bunches than before whose dynamics experience complex collective effects such as intense space charge forces and CSR. FELs and PWFAs require an ability to quickly switch between many different users with various phase space requirements,in practice exotic setups require lengthy tuning. This talk discusses machine learning (ML) and model independent feedback techniques and their application in both the LCLS and European XFEL to maximize the average pulse output energy of FELs by automatically tuning over 100 components simultaneously. We also discuss the creation of non-invasive longitudinal phase space (LPS) diagnostics at PWAs. Finally, we present a hybrid adaptive ML approach in which model-independent methods together with neural networks which has been demonstrated at the LCLS to control electron bunch LPS by tuning FEL components automatically.
Precise accelerator simulations are powerful tools in the design and optimization of exiting and new charged particle accelerators. We all know from experience, the computational burden of precise simulations often limits their use in practice. This becomes a real hurdle when requiring real time computation. I will demonstrate two techniques, based on Polynomial Chaos Expansion and Deep Neural Networks that hints a path forward, towards precise real time computing.
Jefferson Laboratory is a single-purpose, DOE national lab which serves the nuclear physics community through the Continuous Electron Beam Accelerator Facility (CEBAF). This presentation will give a broad view of the ways machine learning (ML) is being used at the lab, with a bias toward applications in accelerator physics. An ML-based system to classify accelerating cavity trips is nearing deployment and some features of the implementation will be highlighted. A few challenges preventing the emergence of a "killer app" will be raised with the aim of promoting discussion during the workshop. Finally, we describe steps being taken at the lab to equip scientists with the skills to leverage machine learning in their areas of expertise and how possible collaboration between laboratories could further enhance proficiency in this field.
In order to improve the performance as well as the experimental setups at the large-scale user facility BESSY II (operated by the Helmholtz-Zentrum Berlin), both the beamline and the machine groups have started working towards setting up the infrastructure to introduce modern analysis, optimization and automation based, among others, on Machine Learning techniques. We introduce some of these first tools concerning data acquisition, simulation, machine measurement prediction and parameter tuning as well as organisational topics such as collaborations and infrastructure.
Imaging atmospheric Cherenkov telescope (IACT) arrays are designed to detect astrophysical sources at very high energies, through image analysis of air showers initiated by gamma rays entering the atmosphere. It is crucial for IACTs to separate the multi-telescope stereo images of gamma-ray signals from the background of charged cosmic-ray particles, the flux of which is several orders of magnitude greater. We use a combination of convolutional neural networks (CNNs) with a recurrent neural network (RNN) to achieve this task in CTLearn, an open source Python package using deep learning to analyze data from IACTs. To allow convolutions of images recorded by some telescopes with a hexagonal pixel layout, we experiment with multiple methods (e.g., oversampling or interpolation). In this workshop, we present the initial performance of the CNN-RNN models and the several methods to process images with hexagonal pixels implemented in CTLearn.
Reliability, availability and maintainability determine whether or not a large-scale accelerator system can be operated in a sustainable, cost-effective manner. The operation of specific accelerator equipment and IT resources requires an increasingly higher focus on data analysis to meet these requirements. In addition setting up the machine for beam can be a time-consuming activity for the operators which can not always be optimised by standard algorithms because of high dimensional data and unclear correlations.
Existing optimiser and ML algorithms can be leveraged to produce models for these problems, based on historical data and/or reinforcement learning. In this presentation we present the fruit of internal discussions at CERN regarding ML applications, such as Linac4 transmission efficiency optimisation using a Powell optimisers and CNNs for classifying beam dump BTVDD images.
The collimation system is installed in the Large Hadron Collider to protect its super-conducting magnets and sensitive equipment from potentially dangerous beam halo particles. The collimator settings are determined following an alignment procedure, whereby collimator jaws are moved towards the beam until a suitable spike pattern, consisting of a sharp rise followed by a slow decay, is observed in nearby beam loss monitors. This indicates the collimator jaw is aligned to the beam. The human task of spike classification can be automated by training a pattern recognition model to automatically classify between alignment and non-alignment spikes. A data set was collated from previous alignment campaigns, from which fourteen manually engineered features were extracted and six machine learning models were trained, analysed in-depth and thoroughly tested. The suitability of using machine learning in LHC operation was confirmed during collimator alignments performed in 2018, which significantly benefited from the models trained through machine learning [1].
[1] G. Azzopardi et al., "Automatic spike detection in beam loss signals for LHC collimator alignment," in Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, vol. 934, pp. 10-18, 2019.
The theory behind various optimization techniques is often not trivial, however their implementation on complex systems such as a Free Electron Lasers could be even more challenging.
In this talk we will present the experience gained using a Machine Learning technique called Reinforcement Learning for the performance optimization of the seeded FEL FERMI.
In particular, the difficulties encountered during its implementation in MATLAB, the obtained results and the possible future developments will be discussed.
The Relativistic Heavy Ion Collider (RHIC) at Brookhaven National Laboratory will undergo a beam energy scan over the next several years. To execute this scan, the transfer line between the Alternating Gradient Synchrotron (AGS) and RHIC or the so-called the ATR line, must be re-tuned for each energy. Control of the ATR line has four primary constraints: match the beam trajectory into RHIC, match the transverse focusing, match the dispersion, and minimize losses. Some of these can be handled independently, for example orbit matching. However, offsets in the beam can affect the transverse beam optics, thereby coupling the dynamics. Furthermore, the introduction of vertical optics increases the possibilities for coupling between transverse planes, and the desire to make the line spin transparent further complicates matters. During this talk, we will explore the use of neural network control policies to perform rapid reconfiguration of the beam-line in simulation. We will begin with an overview of the ATR line and the challenges associated with its tuning. We then discuss the development of our neural network tools and demonstrate their use on simulations of the ATR line.
A 4th generation, 6-GeV high energy synchrotron radiation light source, High Energy Photon Source (HEPS), is being built in suburban Beijing, China. The complexity level and high accuracy requirements for the accelerator and beamlines will need a state-of-art controls approach such as machine learning technology. It is essential to design a modern software architecture with proper modularization to utilize software reusability with sensible API sets. A Python based Machine Learning (ML) Platform for Accelerator High-Level Applications is one of the three planned software platforms. To simplify a converted physicist programmer’s work load, the ML platform provides common functions such as data query, cleaning, graphing and interfaces to popular ML packages like Scikit-Learn and TensorFlow. The data query includes access to EPICS-based archiver systems as well as live data. Possible ideas for ML applications will also be presented.