Electron Neutrino Classification in Liquid Argon Time Projection Chamber Detector
Piotr Płoński, Dorota Stefan, Robert Sulej, Krzysztof Zaremba
EElectron Neutrino Classification in Liquid ArgonTime Projection Chamber Detector
Piotr P(cid:32)lo´nski , Dorota Stefan , Robert Sulej , Krzysztof Zaremba Institute of Radioelectronics, Warsaw University of Technology,Nowowiejska 15/19,00-665 Warsaw, Poland, {pplonski,zaremba}@ire.pw.edu.pl Instituto Nazionale di Fisica Nucleare, Sezione di Milano e Politecnico,Via Celoria 16, I-20133 Milano, Italy [email protected] National Center for Nuclear Research,A. Soltana 7, 05-400 Otwock/Swierk, Poland
Abstract.
Neutrinos are one of the least known elementary particles.The detection of neutrinos is an extremely difficult task since they are af-fected only by weak sub-atomic force or gravity. Therefore large detectorsare constructed to reveal neutrino’s properties. Among them the LiquidArgon Time Projection Chamber (LAr-TPC) detectors provide excel-lent imaging and particle identification ability for studying neutrinos.The computerized methods for automatic reconstruction and identifica-tion of particles are needed to fully exploit the potential of the LAr-TPCtechnique. Herein, the novel method for electron neutrino classificationis presented. The method constructs a feature descriptor from images ofobserved event. It characterizes the signal distribution propagated fromvertex of interest, where the particle interacts with the detector medium.The classifier is learned with a constructed feature descriptor to decidewhether the images represent the electron neutrino or cascade producedby photons. The proposed approach assumes that the position of primaryinteraction vertex is known. The method’s performance in dependency tothe noise in a primary vertex position and deposited energy of particlesis studied.
Keywords:
Electron Neutrino, Classification, Image Descriptor, LiquidArgon, Time Projection Chambers
Neutrinos are one of the fundamental particles as well as one of the least under-stood. They exist in three flavors: electron, muon and tau. There is a hypothesisabout the existence of the fourth type of flavor, namely sterile [15]. The detec-tion of neutrinos is an extremely difficult task since they are affected only byweak sub-atomic force or gravity. Therefore, large detectors are constructed to a r X i v : . [ c s . C V ] M a y P. P(cid:32)lo´nski et al. reveal neutrino’s properties. Among them the Liquid Argon Time ProjectionChamber (LAr-TPC) detector, proposed by C.Rubbia in 1977 [16], provides ex-cellent imaging ability of charged particles, making it ideal for studying neutrinooscillation parameters, sterile neutrinos existence [15], Charge Parity violation,violation of baryonic number conservation and dark matter searches. The LAr-TPC technique is used in several projects around the world [11], [2], [1], [9], [7].Among them, the ICARUS T600 [7] was the largest working detector locatedat Gran Sasso in underground Italian National Laboratory operating on CNGSbeam (CERN Neutrinos to Gran Sasso). In this study, the T600 parameterswill be used since other existing or planned LAr-TPC detectors have the sameor similar construction and settings.A neutrino which passes through the LAr-TPC detector, can interact with nucleiof argon. Density of argon in liquid form makes the rate of interactions prac-tical for experimental study. Charged particles, created in interaction, produceboth scintillation light and ionization electrons along its path in the LAr-TPCdetector. The scintillation light, which is poor compared to ionisation charge, isdetected by photomultipliers which trigger the read-out process. Free electronsfrom ionizing particles drift in a highly purified liquid argon in an uniform elec-tric field toward the anode. Electrons diffussion approximate value 4.8 cm /s ismuch slower than electron drift velocity 1.59 mm/ µ s, therefore they can driftto macroscopic distances preserving high resolution of track details. The anodeconsists of three wire planes, so-called Induction1, Induction2, Collection. A sig-nal is induced in a non-destructive way on the first two wire planes, which arepractically transparent to the drifting electrons. The signal on the third wireplane (Collection) is formed by collecting the ionization charge. The wires inconsecutive planes are oriented in three different degrees with respect to thehorizontal with 3 mm spacing between wires in the plane. This allows to local-ize the signal source in the XZ plane, whereas Y coordinate is calculated fromwire signal timing and electron drift velocity . Signal on wires is amplified anddigitized with 2.5 MHz sampling frequency which results in 0.64 mm spatialresolution along the drift coordinate. The digitized waveforms from consecutivewires placed next to each other form 2D projection images of an event, withresolution 0.64 mm x 3 mm.One of the common aims of proposed and future neutrino experiments is to studythe appearance of electron neutrinos in the muon neutrino beam. Fundamentalrequirement for a such study is the method for classification of the interactingneutrino flavour and, in the case of detectors placed on surface, the method forthe cosmogenic background rejection. Selection of ν e interaction among other ν interactions and background events should involve analysis of the primaryinteraction vertex (PIV) features, including detection of a single electron andpresence of hadronic activity. These features may allow to eliminate the majorityof events that can mimic signal, namely: CERN - European Organization for Nuclear Research Coordinate system labeling is given for reference. e Classification in LAr-TPC Detector 3 – ν interactions with π produced in the vertex, which decays into gammasimmediately, and one or more gammas converts to e+/e- in close vicinity ofthe vertex; – gammas produced by cosmogenic sources, converting in the detector volumewithin the data taking time window with production of e+/e- pair.In this paper, a novel method for automatic classification of ν e from cosmogenicsources is presented. The considered range of energy deposited by an event inthe detector is 0.2-1.0 GeV. We expect that appearance of interesting ν e eventswithin this range and therefore we are preparing the method for rejection of back-ground events resulting in similar energy deposit. In the proposed method foreach event a feature descriptor is constructed. It describes the signal distributionin Induction2 and Collection views. The method assumes that the localization ofthe PIV, where the particle starts interaction with detector, is known. The clas-sifier is learned with a created feature descriptor. Herein, the different settingsused in the feature vector construction are examined. The settings with the bestperformance are selected. The impact of noise in PIV localization on method’sperformance is analyzed. Additionally, the classifier performance is assessed onvarious energy ranges of classified events. The dataset was generated with the FLUKA software [5] and T600 detectorparameters. There were generated 7090 events with energy from 0.2 to 1.0 GeVequally distributed. Among them, there were 3283 events from electron neutrino(positive class label) and 3807 events from cosmogenic sources (negative classlabel). For each event the position of primary interaction vertex is assumed to beknown. All the events have PIV located at least 5 cm from anode or cathode. Allimages were deconvoluted with impulse response of the wire signal readout chain.The segmentation procedure [13] was applied to remove detector’s noise fromimages. There are considered two views for each event, namely, Induction2 andCollection. From each view the image chunk with size 101 wires x 505 samplesand center in the PIV was considered. The used chunk’s size is sufficiently largefor analysis conducted in this paper. The images where downsampled to 101 x 101pixels size to provide similar resolution on both axis, where 1 wire correspondsto 1 pixel in the x-axis and 5 samples corresponds to 1 pixel in the y-axis.
The events from different classes have charge amplitudes propagated in differentways starting from the PIV. Herein, the event’s feature descriptor is proposed
P. P(cid:32)lo´nski et al. to describe this property. The event observed in the detector is described astwo images, from Induction2 and Collection views. Each image is convertedinto the polar coordinate system with radius R and number of bins B spacedto each other with 360 /B degrees and center in the PIV. The total charge ineach bin is summed and creates a charge histogram for the considered view.Additionally, for each view the statistics variables are computed, which describeminimum, maximum, standard deviation, mean and total sum of charge in thehistogram. The histogram values and statistics variables from both views forma feature vector. This results in 2( B + 5) features describing each event. In theFig. 1 are presented images of example events from positive and negative class,with corresponding images after conversion into polar coordinates and chargehistograms. It can be observed that the histograms from the event with negativeclass (Fig.1 c,f) have one peak. Whereas, histograms for positive class (Fig.1i,l) have more than one peak. This is the main difference between positive andnegative classes. What is more, the tracks in images from negative class eventshave broader peaks in the histograms, contrary to tracks from positive classevents which appear as lines in the image and narrow peaks in the histograms.The classifier algorithm is learned to distinguish these properties coded in thefeature vector. In the proposed approach, the created feature descriptor of the event is an in-put for the classifier, which response is a probability whether image representsinteraction of electron neutrino. The Random Forest [4] algorithm was used asclassifier. To asses the classifier’s performance the Receiver Operating Charac-teristic (ROC) [6] curve was used, whereTrue Positive Rate =
T PT P + F N , (1)False Positive Rate =
F PF P + T N . (2)The TP stands for true positives - correctly classified positive samples, TN aretrue negative - properly classified negative samples, FP are false positives -negative samples incorrectly classified, and FN are false negatives, which arepositive samples improperly classified as negatives. Additionally, the Area UnderROC Curve (AUC) and accuracy (the number of all correctly classified samples)was used. They were computed on 5-fold cross-validation (CV) repeated 10 timesfor stability of the obtained results. e Classification in LAr-TPC Detector 5(a) Raw, Ind2, bkg. (b) Polar, Ind2, bkg. (c) Distribution, Ind2, bkg. T o t a l s i gna l (d) Raw, Coll, bkg. (e) Polar, Coll, bkg. (f) Distribution, Coll, bkg. T o t a l s i gna l (g) Raw, Ind2, sig. (h) Polar, Ind2, sig. (i) Distribution, Ind2, sig. T o t a l s i gna l (j) Raw, Coll, sig. (k) Polar, Coll, sig. (l) Distribution, Coll, sig. T o t a l s i gna l Fig. 1: The example of negative (bkg.) and positive (sig.) event presented indifferent perspectives, namely as a raw image observed in the detector, the imagein polar coordinates, and charge distribution with 36 bins and radius equal 10pixels. The each event is presented in Induction2 (Ind2) and Collection (Coll)views. Each row describes one event in a selected view.
P. P(cid:32)lo´nski et al.
The feature vector depends on two parameters: the number of bins and the lengthof the radius. In order to construct the most discriminative feature vector, thevarious parameters combinations were checked. There were considered a num-ber of bins: { , , , } , radius length: { , , , , } pixels and presenceof signal statistics. To asses the discriminative power of feature vector, perfor-mance of the Random Forest (RF) classifier with 1000 trees was measured. Theresults are presented in the Fig.2. There can be observed that performance ofthe classifier is higher when statistics variables are included in the feature vec-tor. The best performance was obtained for the feature vector with 36 bins andradius length 10 pixels and signal statistics included, the AUC is 0.9893 ± ± llllllllll lllll lllll llllllllll lllll lllll Without signal statistics With signal statistics
Number of bins
AUC
Radius lllll
Fig. 2: The performance of the method computed on 5-fold CV repeated 10times for different settings used for event’s feature descriptor construction. Therewere used different number of bins, various length of the radius and presence ofadditional variables with signal statistics.The performance of the proposed method in the dependency to the number oftrees used in the RF classifier is presented in the Fig.3. The AUC of the methodincreases with increasing tree number in the RF. However, the performance for1000 or more trees in the forest is almost stable. In further analysis the 1000trees in the RF were used.The proposed method assumes that PIV position is known. It should be des-ignated before classification by another algorithm, for instance with algorithm e Classification in LAr-TPC Detector 7 presented in [12] and additional logic rules about PIV position. Therefore, theperformance of the method for different noise levels in PIV’s position is exam-ined. The noise levels were generated by drawing a random number of pixels andadding to the true PIV’s location. The ROC curves for different noise levels inPIV are presented in the Fig.4a and the AUC values for 5-fold CV repeated 10times are in Table 1. The performance of the method decreases with more noisein PIV’s location. However, the method’s AUC is 0.9360 for ± l l lll ll ll Number of trees
AUC
Fig. 3: The performance of the method computed on 5 fold CV repeated 10 timesfor different tree numbers for 36 bins and radius length equal 10 pixels and signalstatistics included in feature vector. (a) Dependency to noise in PIV
False positive rate T r ue po s i t i v e r a t e . . . . . . no noise+/− 1 pixel+/− 2 pixels+/− 3 pixels+/− 4 pixels+/− 5 pixels (b) Performance in energy ranges False positive rate T r ue po s i t i v e r a t e . . . . . . Fig. 4: The performance of the proposed method for different (a) noise levels inPIV’s location and (b) different energy of observed event.
P. P(cid:32)lo´nski et al.
The image content depends on the observed event’s energy. Therefore, the methodperformance were tested for different event’s energy ranges (0.2-0.4, 0.4-0.6, 0.6-0.8, 0.8-1.0) GeV. The ROC curves are presented in the Fig.4b and AUC valuesfor 5-fold CV with 10 times repetition are in Table 2. It can be observed thatfor very low energies (0.2-0.4) GeV the performance of the method decreasesslightly. However, for events with energy greater than 0.4 GeV the accuracy ofthe method is almost stable.
Noise level AUCzero noise 0.9893 ± ± ± ± ± ± Table 1: The performance of the proposed method for different noise levels inPIV’s location computed with 5-fold CV repeated 10 times.
Energy range AUC0.2 - 0.4 0.9770 ± ± ± ± Table 2: The performance of the proposed method for different energy of observedevent computed with 5-fold CV repeated 10 times.
The fundamental requirement in the case of surface neutrino detectors is an abil-ity for rejection of cosmogenic background events. Herein, the novel method forclassification of neutrino with electron flavour based on raw images from LAr-TPC detector is presented. The method constructs the feature vector for anobserved event. It describes the distribution of the charge starting from primaryinterest vertex, which position is assumed to be known. The classifier makes adecision whether the detected event is an electron neutrino based on a featuredescriptor. The best combination of parameters used for a feature vector con-struction was selected. The method has AUC 0.9893 and accuracy 0.9535. Theexperiments for different noise levels in PIV locations show that even with largenoise (2 pixels, which corresponds to 2 wires on x-axis and 10 samples on y-axisrandomly added in any direction) the AUC of the method is 0.9360. The perfor-mance of the method is slightly lower for low energy events ( < e Classification in LAr-TPC Detector 9 events with energy higher than 0.4 GeV is almost constant with AUC greaterthan 0.99. The future work will focus on combining the proposed method withalgorithm for PIV position estimation.
PP and KZ acknowledge the support of the National Science Center (Harmonia2012/04/M/ST2/00775). Authors are grateful to the ICARUS Collaboration andPolish Neutrino Group for useful suggestions and constructive discussions duringa preliminary part of this work.