Application of Machine Learning to Beam Diagnostics PDF

11 avr. 2017 Machine learning is a branch of artificial intelligence that has been employed in a variety of applications to analyze complex data sets and ...

Applying Supervised Learning

A supervised learning algorithm takes a known set of input data (the training set) and known responses to the data (output) and trains a model to generate

Application of Machine Learning to Beam Diagnostics

Abstract. Machine Learning (ML) techniques are widely used in science and industry to discover relevant information and make predictions from data.

Application of machine learning models and interpretability

To this end we use a flexible machine learning model

Runtime Verification for Critical Machine Learning Applications

In the last decade the application of Machine Learning (ML) has encountered an increasing interest in various applicative domains

Application of Machine Learning for the IPM-Based Profile

tool and machine learning algorithms lead to a beam profile correction methods for electron-collecting IPMs. INTRODUCTION. Ionization Profile Monitors (IPM)

Multiscale Wavelets on Trees Graphs and High Dimensional Data

Theory and Applications to Semi Supervised Learning. Matan Gavish1 gavish@stanford.edu. Boaz Nadler boaz.nadler@weizmann.ac.il.

Application of Machine Learning To Epileptic Seizure Detection

We present and evaluate a machine learn- ing approach to constructing patient-specific classifiers that detect the onset of an epileptic seizure through

Deep anomaly detection using self-supervised learning: application

10 mars 2022 detection using self-supervised learning: application to time series of cellular data. ASPAI 2021 -. 3rd International Conference on ...

APPLICATION OF MACHINE LEARNING TO BEAM DIAGNOSTICS

E. Fol

1 2, R. Tomás1, G. Franchetti2 3, J. Coello de Portugal4,1CERN, 1211 Geneva 23, Switzerland

2Johann-Wolfgang Goethe University, 60438 Frankfurt am Main, Germany

3GSI Helmholtzzentrum für Schwerionenforschung, 64291, Darmstadt, Germany

4Paul Scherrer Institut, 5232, Villigen, Switzerland

AbstractMachine Learning (ML) techniques are widely used in science and industry to discover relevant information and make predictions from data. The application ranges from face recognition to High Energy Physics experiments. Re- cently, the application of ML has grown also in accelerator physics and in particular in the domain of diagnostics and control. The target of this paper is to provide an overview of ML techniques and to indicate beam diagnostics tasks where ML based solutions can be efficiently applied to com- plement or potentially surpass existing methods. Besides, a short summary of recent works will be given demonstrating the great interest for use of ML concepts in beam diagnos- tics and latest results of incorporating these concepts into accelerator problems, with the focus on beam optics related applications.

MOTIVATION

Traditional optimization tools demonstrate successful per- formance in applications on linear optics corrections and problems with limited amount of optimization targets [1-6]. Bigger challenges emerge when diagnostics of complex non- linear behavior is required and several variables have to be taken into account as final objective. The amount of time and computational power required by traditional methods might become unacceptable for future accelerator facilities. ML is well known for surpassing human performance in some specific tasks such fraud detection, forecasting of mar- ket trends and risks, online recommendations, recognition of voice and images and in general in discovering correlations in large scale data sets. Most of these tasks can find analo- gies in beam control and diagnostics. For example, anomaly detection methods applied for fraud detection can be used to detect defects in the instrumentation and forecasting tech- niques can be transferred to predict beam behavior during operation. Free Electron Lasers (FEL) problems for optimization and diagnostics have to deal with non-linear, multi-objective functions which depend on thousands of time-varying ma- chine components and settings. These properties meet the limitations of traditional optimization methods and make techniques. The main limitation of traditional optimization methods is that the objective function or specific rules and thresholds have to be known. ML methods can learn from given examples without requiring explicit rules. elena.fol@cern.chRELEVANT MACHINE LEARNING

CONCEPTS

ML techniques aim to build computer programs and algo- rithms that automatically improve with experience by learn- ing from examples with respect to some class of task and per- formance measure, without being explicitly programmed [7]. Depending on the problem and existence of learning exam- ples, different approaches are preferred. If pairs of input and desired output are available, an algorithm can generalize the problem from the given examples and produce prediction for unknown input. ML algorithms that learn from input/output pairs are calledsupervised learningalgorithms. Opposite to supervised learning,unsupervised learningalgorithms solve the tasks where only input data is available. Unsupervised learning is suitable for the problems such anomaly detection, signal denoising, pattern recognition, dimensionality reduc- tion and feature extraction. In the following a brief overview on significant machine learning concepts that can be used as supervised as well as unsupervised approaches is presented. We also give a short introduction to Reinforcement Learn- ing - ML technique which recently became of great interest especially for control tasks.

Artificial Neural Network

Artificial Neural Networks (ANNs) are well suited for learning tasks, where data is represented by noisy, complex signals and the target output function may consist of several parameters. A basic ANN consists of a single processing unit (neuron), that takes theweightedinputs and an addi- tional activation function to introduce the nonlinearity in the output. For more complex practical problems, ANNs are composed of several interconnectedhidden layerswith multiple neurons stacked. ANNs can be used for both regres- sion and classification problems. In case of classification the output can be either a class label or a probability of an item belonging to a class. The learning of ANN is per- formed usingbackpropagationalgorithm [8] on a set of examples. For each example the training algorithm com- putes the derivatives of the output function of the network. The obtained gradients with respect to all weights are then used to adjust the weights in order to achieve a better fit to the target output. In backpropagationstochastic gradient optimization method in order to minimize the loss between the network output values and the target values for these outputs by updating the connection weights. ANNs with many hidden layers calleddeep neural networksare able to use fewer neurons per layer and have a better generalization

39th Free Electron Laser Conf. FEL2019, Hamburg, Germany JACoW Publishing

ISBN:978-3-95450-210-3 doi:10.18429/JACoW-FEL2019-WEB03 Electron Diagnostics, Timing, Synchronization, and Controls WEB03

311Content from this work may be used under the terms of the CC BY 3.0 licence (©2019). Any distribution of this work must maintain attribution to the author(s), title of the work, publisher, and DOI

ability [11], however the optimization of the structure and training of these networks is not trivial. There are no strict rules for building ANN architecture (number of neurons, layers, initial weights) as it usually heavily depends on a particular problem. However, techniques to adjust the ar- chitecture parameters exist. A detailed overview on various ANN architectures and training methods and their suitability for different applications can be found in [12-14].

Decision Trees and Ensemble Methods

Decision tree learning is a method for approximating discrete-valued target functions, which are represented by decision trees. Considering the case of classification, deci- sion trees sort down the input instances from the root to leaf nodes. Usually, the splitting is based on one of the input parameters or a specified set of splitting criteria [15,16]. Each leaf corresponds to one class representing the most appropriate class label. For regression problems the leaf nodes correspond to an approximation of target values. and perform poorly on unexplored sample. One possible solution to overcome this problem is to build ensembles of trees [17]. By training several slightly different models and taking the average prediction, the variance of the model can be reduced. ComparedtoANNs, decisiontreesaresimplertointerpret and to understand its way of obtaining the final results and the underlying process, e.g through the feature importance analysis. Featureimportanceanalysishelpstounderstandthe contribution of each input parameter to the decision during the training process. The ability of decision trees to evaluate the importance of input parameter is a significant advantage of these algorithms. Knowing the importance of the features we can reduce the model complexity and simplify the data preprocessing steps without significant accuracy loss.

Clustering

Cluster analysis includes methods of grouping or separat- ing data objects into clusters, such that dissimilarity between the objects within each cluster is smaller than between the objects assigned to different clusters [18,19]. Cluster analy- sis is used in a wide range of applications. Data clusters can be considered as a summarized representation of the data, such that group labels can describe patterns or similarities and differences in the data. Moreover, clustering can be used for prediction, such that classification of unseen data is performed based on knowledge about the properties of present data and by evaluating their similarity to the incom- ing data sample. The significant benefit of cluster analysis is theunsupervised learningapproach, which means that no labeled data is needed to find a solution. The simplest and the most commonly used clustering algorithm is k-means [20], which is based on centroid search. Another kind of clustering algorithms are the density-based algorithms such DBSCAN [21], that views clusters as areas of high density separated by areas of low density, instead of looking for the centroids. Decision tree based methods also can be applied for cluster analysis using the data splits based on different features. Most of cluster analysis techniques allow to build clusters in a multidimensional space. Apart from classification and pattern recognition, clus- ter analysis can be used as denoising method looking for abnormalitiesinthesignal. Moreover, buildingclusterscom- bining a large set of different observables can simplify the data visualization and manual analysis, such elimination of outliers in the measurements and detection of anomalies.

Reinforcement Learning

The concept of Reinforcement Learning (RL) is based on environment-agent interaction [22]. The agent takes an action on the environment, and the environment reacts pro- ducing a reward, which is used by the agent to learn how to improve its actions. The approach does not require an existing data set consisting of input-output pairs, instead the agent is learning based on the continuous interaction with the environment which is varying depending on the action and its own dynamics. Considering this learning principle, RL can be applied to unstable, time-varying problems since the agent should be able to adjust its action to the changes of the response from the environment. The ability of RL makes this approach particularly appealing for the control and optimization of accelerator components. Recent ad- vances on RL application on accelerator control tasks can be found in [23].

OVERVIEW ON CURRENT

APPLICATIONS

In the following we demonstrate some ML applications currently being used in accelerator technology and ongo- ing research on potential ML based approaches. An earlier overview on previous works related to beam diagnostics can be found in [24], for a wider overview on opportunities in

ML for particle accelerators see [25-27].

Virtual Diagnostics

Various instruments and diagnostics techniques are re- quired in order to monitor the beam itself and variables which affect its parameters. Virtual diagnostics can assist in case a direct measurement would have a destructive im- pact on the operation or in the locations where no physical instrumentation can be placed. ML can provide techniques to build such virtual beam diagnostics instruments. Simu- lation studies and experimental demonstrations have been carried out on FACET-II and Linac Coherent Light Source (LCLS) to study ML-based longitudinal phase space (LPS) prediction. Training data for a feed-forward ANN has been acquired from a large number of simulations that represent changes in LPS distribution as response to the change of various accelerator parameters, as well as from the existing measurements at LCLS. ML model demonstrates a good agreement between the prediction and simulated or mea- suredLSPimages[28]. Anotherexampleistheestimationof

39th Free Electron Laser Conf. FEL2019, Hamburg, Germany JACoW Publishing

ISBN:978-3-95450-210-3 doi:10.18429/JACoW-FEL2019-WEB03 WEB03

312Content from this work may be used under the terms of the CC BY 3.0 licence (©2019). Any distribution of this work must maintain attribution to the author(s), title of the work, publisher, and DOIElectron Diagnostics, Timing, Synchronization, and Controls

oscillation amplitude and synchrotron damping time based on LPS measurements at Shanghai Synchrotron Radiation Facility (SSRF) [29]. Here, Gradient descent algorithm is used to estimate the fitting parameters which are then used as target variables in a supervised model. ANN is trained to predict these values from longitudinal phase measurements obtained from the Beam Position Monitors (BPM). Another example from SSRF is a study on correlations between the beam size and the images from multi-slit imaging system, aiming to improve the accuracy of BPMs using ANN [30]. A special kind of ANN,convolutional neural networks (CNN) [31] have been applied at FAST on image based diagnostic during beam operation [32]. A combination of a CNN and a feed-forward NN yields promising results for the prediction of beam parameters on simulated data sets. The the gun phase as inputs and produces a prediction for various downstream beam parameters. Application of ANN can be found also in correction of distorted beam profiled measured at ionization profile monitors (IPM) [33]. ANN model has been trained on IPM simulations in order to establish the mapping between measured profiles together with bunch length and bunch intensity to the original beam profile.

Optimization and Operation

ML methods are especially suited for non-linear and time- varying systems with large parameter spaces. Operation of a complex system such as an accelerator, whose beam dynamics exhibits nonlinear response to machine settings can be considered as a typical ML task. Due to the constant increase of machine design complexity and development of new interacting systems, traditional techniques might be- come insufficient. Reinforcement learning demonstrates aquotesdbs_dbs6.pdfusesText_12

[PDF] Application of Machine Learning to Beam Diagnostics

Machine Learning in Radiology: Applications Beyond Image

Applying Supervised Learning