Reconstruction and Machine Learning in Neutrino Experiments

16 Sept 2019, 08:45 → 18 Sept 2019, 13:00 Europe/Berlin

Building 1b, SR4b (DESY Hamburg)

Björn Wonsak, Caren Hagner

Source: priotography - Flickr: www.flickr.com/photos/priotography/435504045/ - (CC BY 2.0)

The aim of this workshop is to bring together the expertise in reconstruction techniques and machine learning methods collected in the neutrino community. Many of the current and next generation neutrino detectors will have the abilities of fine grained detectors combined with large volumes. This offers unique opportunities but is also a challenge for reconstruction and event analysis. Here we want to collect an overview of modern approaches to this task.

We especially encourage young scientist in presenting their work and using the opportunity to establish valuable contacts.

Notification: A conference fee of 40EU will be collected during the registration in front of the seminar room used. Only cash will be accepted.

 

Slides IMG_20190918_104832.jpg IMG_20190918_104832_small.jpg

Monday 16 September

Registration

Welcome/Introduction

1 Welcome and Introduction

In this talk I will welcome the participants. I will provide all necessary information about the workshop and give our motivations for it.

Speaker: Björn Wonsak (UniHH Opera)

Slides WelcomeTalkRecoMLWorkshop_clean4.pdf

Talks: Reconstruction 1

Convener: Björn Wonsak (UniHH Opera)

2 The Pandora particle flow algorithm

Pattern recognition is an essential stage in the reconstruction of particle interactions in liquid argon time projection chamber detectors, which are used to study the properties of neutrinos. The novel multi-algorithm approach implemented in the Pandora software uses many tens of algorithms to gradually build up an image of the event and has been used successfully for pattern recognition in a number of neutrino physics experiments including MicroBooNE, ProtoDUNE, and the future DUNE experiment. This talk outlines the algorithm flows used by Pandora for reconstructing and identifying neutrino and cosmic ray interactions in these detectors.

Speaker: Mr Andrew Smith (The University of Cambridge)

Slides asmith_pandoraTalk_DESY2019.pdf

3 Reconstructing 3D hit information directly from 2D projections

Many Liquid Argon TPC neutrino detectors record the details of the neutrino interaction in the form of hits projected onto two or three 2D wire views. The goal of reconstruction is ultimately to construct 3D objects compatible with the 2D projections. This talk will present two systems that are able to recover 3D hit information directly from the 2D hits, to serve as a starting point for fully-3D reconstruction. Formally this is an underconstrained inverse problem, akin to deconvolution, and the key to success is the application of an appropriate regularization.

Speaker: Chris Backhouse (UCL)

Slides reco3d_light_theme.pdf reco3d.pdf

10:15 Coffee break

Talks: TRIUMP-Helmholz workshop. Keynote: Perspectives on the Future of Data Intensive Computing

Convener: Oliver Oberst

Talks: Reconstruction 2

Convener: Dr Xin Qian (Brookhaven National Laboratory)

4 Track Reconstruction and Characterization in Liquid Scintillator Detectors at High Energies

Liquid scintillator detectors are well established in the field of neutrino physics (e.g. Borexino, KamLand, JUNO and SNO+). Applications span a wide range from solar neutrino detection over neutrinoless double beta searches to reactor anti-neutrino detection. Most liquid scintillator experiments have implemented algorithms to track high-energetic events like muons. Past analyses have demonstrated that radioisotope production (C-11, Li-9 or He-8) is strongly correlated with muon-induced showers. To identify these events and to reduce the dead-time of the detector, reliable and precise tracking algorithms are a strong necessity for many analyses. This talk will summarize recent developments in the field of muon track reconstruction. Focus will be given to two tracking algorithms ("BackTracking" and "topological track reconstruction"). Different to common track reconstruction approaches, the topological track reconstruction is sensitive to the energy loss along the muon track and allows to resolve topological features like showers which lead to an increase in energy deposition. This might then be used for spatial vetoes.

Speaker: Dr Mikko Meyer (TU Dresden)

Slides meyer_hamburg_final2.pdf meyer_hamburg_final_animations.ppt

5 Distributed imaging for liquid scintillation detectors

Scintillation detectors have been fundamental instruments enabling big discoveries in particle and nuclear physics. Thanks to their wide versatility and relative affordability, they are still nowadays an active area in detector R&D. It is usually assumed that imaging in such a photon-starved and large-emittance regime is not possible. In this talk, I'll go over a novel approach to liquid scintillator that matches appropriate optics with highly segmented photodetector coverage. In particular, if dedicated reconstruction algorithms are employed, this technique can be used to produce images of the radiation-induced events discriminating events produced as a single cluster and those resulting from more delocalized energy depositions. After briefly describing a ML based algorithm used as discriminator, I'll also compare the performances obtained with it with a traditional reconstruction method.

Speaker: Mr Jacopo Dalmasson (Stanford University)

Slides DESY_pres.pdf

6 Event reconstruction in JUNO

The Jiangmen Underground Neutrino Observatory (JUNO) in China is a 20 kton liquid scintillator detector. designed primarily to determine the neutrino mass hierarchy, as well as to study various neutrino physics topics. The large size and the stringent requirement on the unprecedented energy resolution of 3%@1MeV make the event reconstruction in JUNO rather challenging. In this talk, we will focus on the JUNO Central Detector, starting from the PMT waveform reconstruction and then moving on to the current strategies on the vertex and energy reconstruction. Alternative methods will also be mentioned briefly.

Speaker: Dr Wuming Luo (Institute of High Energy Physics, China)

Slides RecAndML_workshop_Hamburg.pdf

13:00 Lunch

At DESY canteen

Talks: Talks: Reconstruction 3

Convener: Abigail Waldron

7 TPC Simulation and Signal Processing for LArTPCs

The single-phase liquid argon time projection chamber (LArTPC) provides a large amount of detailed information in the form of fine-grained drifted ionization charge from particle traces. It has a great capability of identifying various particles and doing energy reconstruction. The Deep Underground Neutrino Experiment (DUNE) which is one of the biggest neutrino experiments in the following 10-20 years will utilize LArTPC technology for a rich assortment of physics. The ongoing Short Baseline Neutrino Program (SBN) also uses three LArTPCs at various baselines to search sterile neutrinos and do other precision physics. In this talk, I’ll focus on a robust signal processing technique that accurately converts the raw digitized TPC waveforms into the number of ionization electrons for induction and collection anode type wire planes. The long-range induction of ionization electrons passing through the wire planes is taken into account and the amplified equivalent noise charge in induction plane is mitigated. This work provides a solid foundation to fully utilize the capabilities of LArTPC and feed to all downstream event reconstruction paradigms. Technical issues and solutions will be discussed. Performance and applications will be shown.

Speaker: Dr Hanyu Wei (Brookhaven National Laboratory)

Slides Diff_RecoWorkshopDESY_LArTPCSPandSim_20190916.pdf

8 Linear system in LArTPC 3D image reconstruction

A liquid argon TPC can be viewed as a linear system where signals of ionization electrons are transformed into measurements of charges on each sensor in a linear and time-invariant fashion. This view allowed us to treat the entire system with linear equations in terms of signal processing and 3D image reconstruction. Techniques used in the Wire-Cell software package will be presented.

Speaker: Dr Chao Zhang (Brookhaven National Laboratory)

Slides Linear_Problem_in_LArTPC.pdf

9 LArTPC Charge-Light Matching and PID with Wire Cell

Wire Cell is a reconstruction package under development for event reconstruction in LArTPC. It consists of many components including i) TPC detector simulation, ii) TPC signal processing, iii) 3D image reconstruction, iv) event clustering, v) charge-light matching, and vi) pattern recognition. In this talk, I am going to describe the Wire-Cell techniques on the Charge-Light matching and the particle identification. The many-to-many charge-light matching is based on the compressed sensing technique. The particle identification includes the track trajectory and dQ/dx fits, which use the advanced techniques in graph theory and linear algebra. The development of these techniques are critical to the overall success of event reconstruction in LArTPC.

Speaker: Dr Xin Qian (Brookhaven National Laboratory)

Slides Application_Wire_Cell.pptx

16:00 DESY tour

Please subscribe for this event during the registration.

Tuesday 17 September

Talks: TRIUMPF-Helmholz Workshop. Keynote: How to know, where to look - prioritising in computer vision

Convener: Simone FRINTROP

Talks: Talks: Machine Learning 1

Convener: Fatih Bay

10 Machine Learning techniques on NOvA

NOvA is a long baseline neutrino oscillation experiment. The NOvA experiment has made measurements using the disappearance of muon and the appearance of electron neutrinos and anti-neutrinos in the NuMI beam at Fermilab including the neutrino mass hierarchy and the lepton CP violating phase. Key to these measurements is the application of machine learning methods for identification of neutrino flavor. The use of these tools, adapted from computer vision, is becoming more widespread within NOvA and the field. These methods require rigorous validation to both understand and develop. I will present an overview of the NOvA experiment and machine learning techniques used for event selection as well as validation techniques used for these algorithms.

Speaker: Micah Groh (Indiana University)

Slides presentation.pdf

11 Particle Identification Using Convolutional Neural Networks in the NOvA Experiment

In 2016, NOvA was the first HEP experiment to employ a convolutional neural network (CNN) in a physics result, using the CNN to classify neutrino events. The physics analyses performed by NOvA can be improved by further identification and reconstruction of particles in the interaction final states. We have developed the first implementation of a CNN for single particle classification which employs context-enhanced inputs. Using contextual information from the neutrino interaction that produces the particles provides additional information to the training, extending the capabilities of our original classifier. This implementation uses a four-tower siamese architecture for separation of independent inputs and inclusion of contextual information. This classifier distinguishes between electrons, muons, photons, pions, and protons with a global efficiency and purity of 83.7% and 83.5%, respectively. In this talk I will describe our implementation of NOvA's single particle CNN, discuss the advantages of adding context information, provide case-studies of the applications and planned future improvements to the classifier.

Speaker: Mr Ryan Murphy (Indiana University)

Slides ProngCNN_RecoAndMLGermany2019.pdf

10:45 Coffee break

Talks: Talks> Machine Learning 2

Convener: Dr Oliver Schulz (MPI Munich)

12 Machine Lerning Applications in Minerva

MINERvA is an experiment exploiting the NuMI neutrino beam line at Fermilab to perform high-precision measurements of neutrino-nucleus interactions on a wide variety of nuclei. A precise understanding of the neutrino-nucleus cross sections is important to reduce systematic uncertainties in the determination of neutrino oscillation parameters and to discriminate between the plethora of nuclear models. In order to improve the measurement of neutrino-nucleus cross section measurement, it is important to precise understanding of the final state particle topology, event selection and so on which in turn demands the application of advance algorithms to maximize the physics output. I will present novel applications of Machine Learning based techniques facilitating the interaction vertex reconstruction, neutral pion reconstruction and hadron multiplicity in the final state.

Speakers: Dr Anushree GHOSH (Universidad Tecnica Federico Santa Maria, Chile), Dr Anushree Ghosh (UTFSM, Chile)

Slides Talk_Desy_ML.pdf

13 Event Classification using Neural Networks in DUNE

The Deep Underground Neutrino Experiment is a next-generation neutrino oscillation experiment that aims to measure CP-violation in the neutrino sector. A deep learning approach based around a convolutional neural network has been developed to provide highly accurate and efficient selections of electron neutrino and muon neutrino interactions. The electron neutrino (antineutrino) selection efficiency peaks at 90% (94%) and exceeds 85% (90%) for reconstructed neutrino energies between the 2-5 GeV. The muon neutrino (antineutrino) event selection is found to have a maximum efficiency of 96% (97%) and exceeds 90% (95%) efficiency for reconstructed neutrino energies above 2 GeV. These event selections are critical to maximise the sensitivity of the experiment to CP-violating effects.

Speaker: Saul Alonso-Monsalve (CERN)

Slides sam_ml_hamburg.pptx

14 Regression CNN Based Energy and Vertex Reconstruction at DUNE

DUNE is the next-generation flagship neutrino experiment designed to decisively measure neutrino CP violation and mass hierarchy. DUNE far detectors are based on liquid argon time projection chamber (LArTPC) technology, which offers an excellent spatial resolution, high neutrino detection efficiency, and superb background rejection. Reconstruction of neutrino events in DUNE's high-resolution detectors is challenging. It is complicated by missing energy due to argon impurity, nonlinear detector energy responses, invisible energy, hadron identities (mass), and overlaps between lepton and hadron interactions. To address these issues, neutrino events can be reconstructed directly from pixel map images of interactions in DUNE's detectors with deep learning methods - in particular, Convolutional Neural Networks (CNNs). In this talk, we will focus on recently developed regression CNNs to reconstruct neutrino energy and interaction vertices. Compared with traditional reconstruction, these methods show a significantly better performance in Monte Carlo simulation.

Speaker: Prof. Jianming Bian (University of California, Irvine)

Slides DUNEDLReco_HAM_09_17_2019.pdf

12:50 Lunch

Talks: Talks: Machine Learning 3

Convener: Prof. Jianming Bian (University of California, Irvine)

15 Machine Learning and Computer Vision for Particle Imaging Detectors in Neutrino Experiments

With firm evidence of neutrino oscillation and measurements of mixing parameters, neutrino experiments are entering the high precision measurement era. The detector is becoming larger and denser to gain high statistics of measurements, and detector technologies evolve toward particle imaging, essentially a hi-resolution "camera", in order to capture every detail of particles produced in a neutrino interaction. The forefront of such detector technologies is a Liquid Argon Time Projection Chamber (LArTPC), which is capable of recording images of charged particle tracks with breathtaking resolution. Such detailed information will allow LArTPCs to perform accurate particle identification and calorimetry, making it the detector of choice for many current and future neutrino experiments. However, analyzing hi-resolution imaging data can be challenging, requiring the development of many algorithms to identify and assemble features of the events in order to reconstruct neutrino interactions. In the recent years, we have been investigating a new approach using deep neural networks (DNNs), a modern solution to a pattern recognition for image-like data in the field of Computer Vision. A modern DNN can be applied for various types of problems such as data reconstruction tasks including interaction vertex finding, pixel clustering, and particle/topology type identification. In this talk I will discuss the challenges of data reconstruction for imaging detectors, recent work and future plans for developing a full LArTPC data reconstruction chain using DNNs.

Speaker: Dr KAZUHIRO TERAO (SLAC National Accelerator Laboratory)

Slides 2019-09-17-DESY-NuMLRecoWorkshop(3).pdf

16 Machine learning applications for JUNO

The new generation neutrino experiment JUNO is a multipurpose experiment with the main goals of determining the hierarchy of neutrino masses and precisely measuring the neutrino oscillation parameters. Thanks to the huge mass of the liquid scintillator target that is equal to 20 kt and 18000 large 20-inch PMTs and 25000 small 3-inch PMTs the detector will collect hundreds of millions of events originating from various interactions and detailed information about each of them. Machine learning methods can provide an effective alternative to the common-used numerical and analytical approaches for event reconstruction, particle identification and candidate selection. Different compositions of the Convolutional Neural Networks (CNN) and the Full Connected Neural Networks (FCNN) are applied for particle identification and muon reconstruction as well as for energy, position and waveform reconstructions of the event. The Long Short-Term Memory neural network (LSTM) which is one of the types of the Recurrent Neural Networks (RNN) may be used for waveform reconstruction.

Speaker: Dr Maxim Gromov (SINP MSU, JINR)

Slides Gromov_ML_Hamburg.pdf

17 Machine Learning methods in Borexino Experiment

The Borexino detector is a liquid scintillator detector located in the Laboratori Nazionali del Gran Sasso (LNGS), in the mountains of central Italy, aiming to measure the low-energy solar neutrinos. It is equipped with nominally 2212 photomultpliers (PMTs), detecting the arrival time of the light produced by the events. In this talk we study the pulse shape discrimination and vertex reconstruction with the machine learning method based on Borexino data.

Speaker: Mr Yu Xu (IKP2 FZJ)

Slides desy_ML.pdf

18 Applications of the Topological Track Reconstruction to low energy events

This contribution presents the application of the Topological Track Reconstruction (TTR) to low energy events of a few MeV and its capability for particle discrimination. In liquid scintillator detectors like the Jiangmen Underground Neutrino Observatory (JUNO) the low energy regime consists of electrons, positrons, or gammas, which are treated as point-like. While positrons and electrons travel typically a few cm, the mean range of gammas can be up to 50cm. This yields the possibility to discriminate electrons and gammas directly via their travel range. Since positrons annihilate with the emission of two gammas the range can also be used to distinguish between electrons and positrons. The discrimination of electrons and positrons gives the ability to suppress the background originating from β+ decays of cosmogenic isotopes in measuring solar 8B neutrinos, whereas the distinction between electrons and gammas comes in handy to suppress natural radioactivity in the detector. The TTR is capable with both conventional and Machine Learning techniques to provide e+/e− and γ/e− discrimination. This makes the TTR an important tool for background suppression. This talk covers an introduction to the TTR, a motivation for particle discrimination, the used techniques and their results. In addition it gives an outlook for the application of the TTR to water-based liquid scintillator detectors.

Speaker: Malte Stender (University Hamburg - Institute for experimental physics)

Slides TTRtoLowEnergyEvents

16:00 Coffee break

Talks: Talks: Machinhe Learning 4

Convener: Joshua Renner (Instituto Galego de Fisica de Altas Enerxias (IGFAE) - University of Santiago de Compostela)

19 Event identification in the NEXT experiment using CNNs

NEXT (Neutrino Experiment with a Xenon TPC) is a neutrinoless double-beta decay experiment that is currently operating a 5 kg-scale demonstrator at the Canfranc Underground Laboratory (LSC). In order to detect such rare events an optimal background identification is necessary. As neutrinoless double-beta events will have a fixed energy, background events can be rejected via energy selection, but for the events that fall in the energy window of the expected signal the two types of events differ only in the topological signature that the particle leaves inside the chamber - i.e., the shape of the track. Since the latter is a special case of 3D Computer Vision applications, the Convolutional Neural Network (CNN) can be used to classify signal vs background. In this talk I will demonstrate how such networks can be trained and applied in a proof-of-concept analysis where the electron positron pair annihilation, having similar topological signature as double electrons from the double-beta decay, is used to mimic the signal.

Speaker: Dr Marija Kekic (IFIC)

Slides DESY_mkekic.pdf

20 Deep Learning in the EXO-200 experiment

The EXO-200 experiment searches for the neutrinoless double beta (0νββ) decay in 136Xe with an ultra-low background single-phase time projection chamber (TPC) filled with 175kg isotopically enriched liquid xenon (LXe). The detector has demonstrated good energy resolution and background rejection capabilities by simultaneously collecting scintillation light and ionization charge from the LXe and by a multi-parameter analysis. The combination of both signatures allows for complementary energy estimates and for a full 3D position reconstruction. This talk will briefly present the concept of the detector and summarize the potential of Deep Learning based methods toward improving low-level event reconstruction and high-level data analysis in the EXO-200 experiment.

Speaker: Mr Tobias Ziegler (Erlangen Centre for Astroparticle Physics)

Slides Presentation.pdf

21 ML techniques for the GERDA 0nbb experiment and beyond

The GERDA experiment is designed to search for neutrinoless double beta decay of Ge-76, using an array of isotopically enriched high-purity germanium detectors. While GERDA is built using state-of-the-art low-background construction techniques, these alone cannot provide the ultra-low background levels required for and achieved by the experiment. We present the data analysis and machine learning (ML) techniques currently employed by the GERDA collaboration to substantially reduce background levels during offline data processing. We also present promising novel approaches for future improvements that may benefit both GERDA and next-generation Ge-76 double beta decay experiments.

Speaker: Dr Oliver Schulz (MPI Munich)

Slides oschulz-gerda-ml-confrec.pdf

Social Dinner

Wednesday 18 September

Talks: Talks: Machine Learning 5

Convener: Viktor Pec

22 Deep Learning in KM3NeT

The KM3NeT research infrastructure will comprise two large underwater Cherenkov neutrino telescopes in the Mediterranean sea. The ARCA detector will be located 100 km off-shore Capo Passero (Italy); its main science objective is the detection of high energy cosmic neutrinos. The ORCA detector will be located 40 km off-shore Toulon (France); its goal is the determination of the neutrino mass hierarchy. Deep convolutional neural networks are being used to reconstruct the event properties, including direction and energy of the incident particle, vertex position and event topology – a proxy for the flavor of the interacting neutrino. Background/signal separation and uncertainty estimation are provided as well. The main purpose of this contribution is to show the results achieved by deep convolutional neural networks and to compare their outputs with conventional reconstruction algorithms. The comparison shows that these first deep learning models already yield results and performance competitive with the official reconstruction pipeline.

Speaker: Simona Maria Stellacci (INFN)

Slides Deep_Learning_in_KM3NeT.pdf

23 Measurement of Neutrino Energy Spectra via Deconvolution with IceCube

IceCube is a cubic kilometer neutrino detector located at the geographic South Pole. Neutrino events in IceCube are detected via Cherenkov light emitted by charged leptons, produced in charged- or neutral current interactions with nuclei in the ice or the bedrock. As these interactions are stochastic processes, the energy of the neutrino cannot be accessed directly, but has to be inferred using the reconstructed energy of the leptons or other energy estimators. Additional smearing effects, introduced by particle propagation and the detector itself, further complicate the problem. This talk will discuss spectral measurements by IceCube obtained using various deconvolution/unfolding analyses as well as their development over time.

Speaker: Dr Tim Ruhe (TU Dortmund)

Slides IceCube_ML_and_Unfolding.pdf

24 Image Based Reconstruction and Deep Learning in MicroBooNE

Operating a 170 ton LArTPC 470 meters from the Booster Neutrino Beam (BNB) at Fermilab, MicroBooNE is applying parallel analysis paths towards the excess of low energy νe-like (LEE) events observed by MiniBooNE. Having been recorded with a high resolution, LArTPC data translated into fine 2D images is ideal for applying computer vision and machine learning techniques. MicroBooNE has applied OpenCV, semantic segmentation network (SSNet) and mutli-particle identification (MPID) network to event reconstruction and selection. In this talk, I will present the image based event reconstruction and deep learning tools used in MicroBooNE.

Speaker: Rui An (Illinois Institute of Technology)

Slides DESY2019_uboone.pdf

10:50 Coffee break

Talks: Talks: Machine Learning 6

Convener: Dr Daniel Bick (Universität Hamburg)

25 Particle tracking based on fully unsupervised disentanglement of the geometrical factors of variation

Particle tracking detectors allow studying of elementary particle interactions and precise measurement of their properties by observing their tracks. Robust tracking algorithms is nowadays a fundamental component of all tracking detectors. Conventional track data reconstruction is performed in three steps. First, 3D tomographic images of the emulsion detector are acquired using automated scanning microscopes. Next, the position of silver grains (“hits”) is located in the 3D image volumes, and finally tracks in the detector volume are reconstructed as sequence of hits.Conventional track data reconstruction is performed in three steps. First, 3D tomographic images of the emulsion detector are acquired using automated scanning microscopes. Next, the position of silver grains (“hits”) is located in the 3D image volumes, and finally tracks in the detector volume are reconstructed as sequence of hits. Several tracking algorithms were developed over the course of evolution of the scanning systems, allowing for efficient track reconstruction in real-time during the acquisition. While satisfying need of many experiments they have several drawbacks. Adaptation to different experimental condition, e.g. high track density or high background level requires tedious calibration ranging from extensive parameter tuning to performing dedicated test runs using e.g. accelerator beams. In addition, when the procedure of extracting the hits is separated from track reconstruction, the tracking algorithm cannot fully exploit the information available in the raw image data, compromising performance especially in the high background/track density cases. Incorporating tracking based on the classical Deep Learning, where the track parameters are predicted from the raw image data would naturally address the latter issue. Yet, to train such model in a supervised manner either one would need to provide massive amount of labeled 3D raw image data for each experimental case, or training would need to be performed largely on the simulated datasets. Training such models in an unsupervised manner, i.e. where no track parameters (labels) are to be provided during the training can address the mentioned issues simultaneously, by both leveraging raw image data for efficient track reconstruction and would allow simple adaptation to new configurations, requiring the raw image dataset only. Here we introduce a tracking approach based on the Deep Convolutional Autoencoder model that learns to disentangle the factors of variation in a geometrically meaningful way in a fully unsupervised manner by imposing equivariance of the space transformation. While the reconstruction constraint alone fails to disentangle the factors of variation in a meaningful way, we show that adding a simple constraint on translational invariance along the track line does not lead to an improvement. We demonstrate that incorporating more sophisticated transformations in the latent representation is demanded to avoid the reference ambiguity.

Speaker: Dr Mykhailo Vladymyrov (University of Bern)

Slides MVladymyrov_Talk_18.09.2019.pdf

26 Machine learning in PET imaging with PETALO

PETALO (Positron Emission Time-of-flight Apparatus with Liquid xenOn) is a novel PET imaging detector concept in which liquid xenon (LXe) scintillating cells are arranged in a ring around the observed volume, and silicon photomultipliers (SiPMs) lining the cells detect the primary scintillation created by 511 keV gamma rays emitted in positron-electron annihilation within the volume. The fast scintillation decay time of LXe could allow for a high-resolution determination of the location of emission points through time-of-flight (TOF) analysis. However, the reconstruction of gamma ray interactions within the LXe cells presents several challenges. In particular, a large fraction of such gammas interact by Compton scattering and leave multiple distinct regions of energy deposition in a cell. We discuss the use of a neural network in identifying such events based on the scintillation pattern left on the SiPMs, and we consider the direct integration of this network in the PETALO electronic readout system to quickly tag and/or reject events that may pose potential problems in reconstruction.

Speaker: Joshua Renner (Instituto Galego de Fisica de Altas Enerxias (IGFAE) - University of Santiago de Compostela)

Slides renner_machine_learning_PETALO_Hamburg.pdf

27 Closing Words, maybe more

Speaker: Björn Wonsak (UniHH Opera)