Particle Identification at MeV Energies in JUNO
PPrepared for submission to JINST
Particle Identification at MeV Energies in JUNO
L. Ludhova b , c H. Rebber a , B. Wonsak a Y. Xu b , c , a University of Hamburg, Institute of Experimental Physics,Luruper Chaussee 149, 22761 Hamburg, Germany b Forschungszentrum Jülich IKP,Wilhelm-Johnen-Strasse, D-52428 Jülich, Germany c III. Physikalisches Institut B, RWTH Aachen University,Aachen, Germany
E-mail: [email protected] , [email protected] Abstract: JUNO is a multi-purpose neutrino experiment currently under construction in Jiangmen,China. It is primary aiming to determine the neutrino mass ordering. Moreover, its 20 kt target massmakes it an ideal detector to study neutrinos from various sources, including nuclear reactors, theEarth and its atmosphere, the Sun, and even supernovae. Due to the small cross section of neutrinointeractions, the event rate of neutrino experiments is limited. In order to maximize the signal-to-noise ratio, it is extremely important to control the background levels. In this paper we discuss thepotential of particle identification in a large liquid scintillator detector like JUNO. We discuss theunderlying principles of particle identification and its application in the experiment. In order toinvestigate the potential of event discrimination, several event pairings are analysed, i.e. α / β , p / β , e + / e − , and e − / γ . We compare the discrimination performance of advanced analytical techniquesbased on neural networks and on the topological event reconstruction keeping the standard Gattifilter as a reference. We use the Monte Carlo samples generated in the physically motivated energyintervals. We study the dependence of our cuts on energy, radial position, PMT time resolution,and dark noise. The results show an excellent performance for α / β and p / β with the Gatti methodand the neural network. Furthermore, e + / e − and e − / γ can partly be distinguished by means ofneural network and topological reconstruction on a statistical basis. Especially in the latter case,the topological method proved very successful.Keywords: Data processing methods; Liquid scintillator detectors; Neutrino detectors; Largedetector systems for particle and astroparticle physics; Particle identification; Machine learning;Topological reconstruction. Corresponding authors. a r X i v : . [ phy s i c s . i n s - d e t ] J u l ontents α / β and p / β Discrimination 124.2 e + / e − Discrimination 154.3 e − / γ Discrimination 18 α / β discrimination 22B Full collection of plots for p / β discrimination 24C Full collection of plots for e + / e − discrimination 26D Full collection of plots for e − / γ discrimination 28 Liquid scintillator (LS) technology has a key role in the detection of low energy neutrinos. Thealmost linear relation between energy deposition and light emission enables calorimetric measure-ments even in the sub-MeV regime. Present and future experiments instrument large target massesin the order of kilotons in unsegmented tanks in order to address the unsolved issues in neutrinophysics, which include the neutrino mass ordering [1], CP violation in neutrino oscillations [2, 3],and neutrinoless double beta decay [4, 5]. Furthermore, the determination of low energy neutrinofluxes offers a unique way to study energy production in the Earth and Sun, as well as the dynamicsof supernovae.Various channels enable neutrino detection in a LS detector, e.g.• inverse beta decay (IBD): ¯ ν e + p → e + + n ,• elastic scattering (ES) with electrons: ν + e → ν + e ,– 1 – elastic scattering with protons: ν + p → ν + p .Identifying signal events is crucial to all neutrino studies, since background is usually dominatingthe event rates. The IBD channel features a characteristic coincidence signature which arisesfrom a delayed gamma emission following the n -capture and resulting in light emission ∼ µ safter the prompt e + signal. Although most backgrounds can be suppressed by time and spacecoincidence requirements, the n -accompanied β − decay of cosmogenic isotopes mimics the signalpattern. The ES channels on the other hand cause single energy depositions and thus cannot bedistinguished from point-like background events due to, for example, radioactive contaminants ofthe construction materials. Usually, an optimization of the signal-to-background ratio can only beachieved by selecting a proper energy range as well as the so called fiducial volume, a wall-lessregion of the LS, defined through the reconstructed vertex.The identification of particle type (PID) is an appealing concept since it offers an independentway for background reduction and can hence lead to an enhancement of the detector sensitivity.Characteristic decay sequences and event topologies, as well as differences in the scintillationprocesses, can affect the topology of the emitted light, and thus, provide handles for PID. Pulseshape analysis, investigating the temporal and/or spatial distribution of detected photons, has provento be a powerful discrimination tool in several neutrino LS experiments. Borexino established areliable discrimination of α and β [6] and also a statistical discrimination of e + and e − events [7].A similar approach was followed by Double Chooz [8]. Double Chooz used PID also for thediscrimination of e + and protons so as to reject fast neutron background from IBD samples [9]. Adiscrimination between neutron and gamma events was successfully studied for the proposed LENAexperiment [10].This work focuses on four general discrimination categories: α / β , p / β , e + / e − , and e − / γ .Our studies are based on MC simulations for the upcoming JUNO detector, making use of itsextraordinarily high yield of detected photons of ∼ The Jiangmen Underground Neutrino Observatory (JUNO) [1] is a next generation neutrino ex-periment currently being built ∼
680 m underground in Jiangmen in south China. Its large targetmass and excellent energy resolution offer exciting opportunities for research in neutrino physics.Figure 1 shows a schematic view on the setup. The heart of the experiment is the central detector(CD), an acrylic sphere with a radius of 17.7 m holding 20 kt of LAB-based LS with admixtures of– 2 – igure 1 . Schematic view of the JUNO detector.
PPO and Bis-MSB. The characteristic length for light attenuation exceeds 20 m in order to makeup for the huge CD dimensions. For light detection, ∼ ∼ ∼ ∼ R .Light attenuation has the strongest effect in the detector center, from where the curve rises towardsthe detector edge, whereas total reflection diminishes the yield in the outermost region above thepeak observed around R =
16 m. The overall high p.e. statistics result in an unprecedented energyresolution for large LS detectors of 3 % / (cid:112) energy / MeV.Background in JUNO is mostly assigned to one of the four main categories: internal background due to the decays of radioactive contaminants of LS (natural U and Th chains,
Pb/
Bi,
Po, C, Kr), external background from gammas penetrating inside the LS volume from outside (e.g.– 3 – igure 2 . Relative number of registered photons as a function of detector volume. from stainless steel frame, PMTs, acryllic vessel), cosmogenic background induced by cosmic muoninteractions, and in some particular cases, neutrino events from other than the envisaged sources.In the following, we give an overview on JUNO’s various physics goals and the respective mainbackgrounds. Especially, the potential contributions to background reduction with PID are pointedout. • The main purpose of JUNO is the determination of neutrino mass ordering (MO). Electronantineutrinos from two nuclear power plants will be detected via IBD with visible energiesfrom the e + component ranging from 1 MeV to 10 MeV. Given the good energy resolution,the MO follows from the measurement of a subdominant oscillation imprinted on the e + energy spectrum. With rates comparable to the MO signal, the most serious background isexpected from the cosmogenic spallation products He and Li, both having the potential toundergo ( β − + n ) decays. The resulting signals coincide temporally and spatially and hence,mimic the IBD signature. Muon vetoes can suppress such events by the cost of roughly 15%exposure loss [1]. An e + / e − discrimination, even if not feasible on event-by-event basis,would mean valuable input for the direct measurement of He and Li production rates. Thelatter has been carried out in KamLAND [12], Borexino [13], Daya Bay [14], and DoubleChooz [15]. For He, only KamLAND could measure a rough yield value, while DoubleChooz and Borexino provided upper limits [15]. JUNO would be able to use the combinedpotential of very large exposure and a statistical PID. The results can in turn find use in theoptimised design of more efficient muon vetoes.• Although suffering from an overburden which is relatively low compared to other undergroundexperiments such as LS-based Borexino [16] or water-based SuperKamiokande [17], JUNOcan contribute to solar neutrino measurements. Combining the large volume and highlight yield, it has a large potential to observe the B solar neutrinos with decreased energythreshold and high statistics. The measurement of other solar neutrino species below 2 MeV( Be, pep , pp ) will strongly depend on the internal contamination of the LS. Since the– 4 –ignal is given by ES with target electrons, all kinds of single events in the energy rangeof interest represent background. Besides internal radioactivity, the external backgrounddemands a fiducial volume cut several meters deep into the CD sphere. A key measurementwill be of B neutrinos down to ∼ C and from the γ ’s of external background and from neutron captures [20].The e − / γ discrimination has thus a potential to expand the exposure significantly, althoughhigh reliability is required due to the exponentially growing rate of γ events towards the CDedge. C undergoes β + -decay, followed after 1 ns by a 718 keV γ -transition, and is reducibleby e + / e − discrimination.• Geo-neutrinos , ¯ ν e created in natural β − decays inside the Earth’s crust and mantle, are aunique tool to asses the Earth’s radiogenic heat, a key parameter to global understanding ofour planet. With roughly 400 events per year, JUNO is expected to collect the world’s largestsample of geo-neutrinos within one year of measurement. Since the detection channel is IBDwith e + signals below 3 MeV, reactor antineutrinos are inevitable background. Cosmogenic He and Li contribute as discussed above. Furthermore, it is known from the experience inKamLAND [21] that C ( α, n ) O reactions constitute another background component dueto various ways to create a prompt signal, one of which is neutron elastic scattering on aproton [22]. β / p discrimination can reject such events.• The rare occurrence of a core-collapse supernova in our galaxy would flush JUNO with allkinds of neutrinos and antineutrinos, triggering a whole bunch of detection channels. Amongthese, IBDs will make up the highest signal rate, exceeding by far the rates from reactorantineutrinos and associated backgrounds. One particular channel open to all neutrinospecies is given by the ES off protons. Being singles with visible energies mainly below1 MeV, the signals are hard to distinguish from radioactivity background. Main contributionscome from the β − emitters Kr and
Bi, and below 0.2 MeV especially from C. All couldbe rejected with e / p discrimination. PID would further help to distinguish the signal fromsupernova channels like electron ES.• The diffuse supernova neutrino background (DSNB), a low isotropic flux of neutrinos ex-pected from the cumulated supernova rate in our universe, has never been detected, yet.With its large target mass, JUNO could find between one and two DSNB events per yearas IBDs [23]. High event rates of reactor antineutrinos rule out detections below 10 MeV,while charged current interactions of atmospheric neutrinos start to dominate over DSNBabove 30 MeV. In between, LENA studies [24] show that the remaining backgrounds from fastneutrons and neutral current interactions with atmospheric neutrinos can be reduced belowthe expected signal level with the help of PID. The signal formation in LS detectors is mainly induced by ionising particles causing an excitationof LS molecules along their path. The subsequent de-excitation goes along with isotropic light– 5 –article type Fast Intermediate Slow τ / w τ / w τ / w [ns] / [%] [ns] / [%] [ns] / [%] γ , e + , e − α Table 1 . Time constants τ i and relative weights w i assumed for the three exponential contributions to thelight emission curves (Eq. 3.1) for different particle types assumed in the JUNO MC simulation. emission in the shortwave part of the optical spectrum. Several particle-related effects alter thedetected pulse shape and can be exploited for PID. In the case of α / β and p / β discriminations,the most striking difference can be traced back to the time curves representing the emission ofscintillation photons. The general behaviour of this process can be described by a superposition oftypically n = φ em ( t ) = n (cid:213) i = w i τ i e − t − t τ i with n (cid:213) i = w i = , (3.1)each parametrised with a weight w i and time constant τ i . The de-excitation of singlet states leads to adominant fast emission component. Excited triplet states lose their energy mainly via non-radiativeprocesses rather than light emission [25–27]. However, interactions with excited triplet states maycreate further excited singlet states which then decay, leading to a suppressed and hence sloweremission component. Furthermore, the interaction of excited singlet states with each other favoursionisation quenching [28], i.e. the light yield per unit of deposited energy is being reduced. Asa consequence, the ratio between the emission components depends strongly on the concentrationof excited states. Particles like α and p entail higher ionisation rates along their path and causemore quenching compared to e + , e − , and γ . Accordingly, the w i and τ i take characteristic valuesfor certain groups of ionising primary particles as can be seen in Table 1, which lists correspondingvalues expected for the JUNO LS mixture. The effect on the pulse shape is compared exemplarilyfor α and β particles in Fig. 3 (a). Both hit time profiles are constructed as a superposition of 1000MC simulated events with visible energies between 0.2 MeV and 1.5 MeV and disregarding TTS ofPMTs. The hit times were corrected by the photon time of flight (ToF) between the vertex point andPMT. One observes considerably higher expectation for late emission times ( >
200 ns) for α events.We note, that the time profiles for protons are expected to be very similar to those of α ’s (Table 1).Regarding e + and e − , weights and time constants are almost identical. PID is instead based onprocesses including positronium formation and positron annihilation. While the ionisation lossesper track length are almost equal for both particles, the e + will most probably form a short-livedmeta-state with a local electron called positronium (Ps) before finally annihilating into two 511 keV γ . Depending on the spin configuration, the decay time in LS is either 125 ps (para-Ps) or 3 ns(ortho-Ps). The fraction of ortho-Ps formation was reported to lie around 50% [29–31]. Since thedecay time for ortho-Ps is comparable to the dominating fast time constant for scintillation (4.93 ns)and to the time resolution of the PMTs, its pulse shape, influenced by the emission of delayedannihilation photons, could be recognized. Moreover, e + events feature a characteristic topology:– 6 – a) (b) Figure 3 . The MC-based time profiles of light emission expected for different particles in JUNO, based onthe parameters from Table 1. Comparison of (a): α and β and (b): e + and e − time profiles. in contrast to the e ± track, which ends after a few cm for kinetic energies below 10 MeV, the γ particles typically undergo several Compton scattering processes, each of which with a mean freepath of tens of cm. Since optical photons travel ∼
20 cm per ns, the spread of e + topology shouldalso leave tiny detectable traces on the pulse shape compared to a point-like e − event. Figure 3 (b)displays how both effects slightly shift the peak position of the e + time profile to higher times. Thedepicted profiles were generated like in (a) with visible energies ranging from 1 MeV to 10 MeV.Ortho-Ps was considered (see Sec. 3.3).With the latter argument, also the direct e − / γ discrimination comes into reach. However,here the extended topology of the gamma events only stems from the number Compton scatteringprocesses needed to release the ganma energy into the scintillator and can not profit from thepresents of a primary particle in addition to the gammas (as is the case for the positrons).Although all PID methods introduced in the following are based on the hit times measured bythe PMTs, they fall into two categories. Firstly, the pulse shape is evaluated directly from the hittimes measured by the PMTs. We refer to them as direct methods. Secondly, the hit times are usedto create a topological event map prior to further analyses. As discussed above, one can discriminate two event classes based on their characteristic timeprofiles. Accordingly, it is required to know the vertex point of the event in order to do a ToF-correction. For the actual discrimination, Gatti filters [32] are commonly used, e.g. in Borexino [7].In the Gatti analysis, it is required to know the expected time profile P i ( t ) for each particle i . Theprofiles serve as density distributions of the probability r i ( t n ) for a particle i to register a PMT hitbetween two times t n and t n + as of r i ( t n ) = ∫ t n + t n P i ( t ) dt . (3.2)– 7 –iven the binned pulse shape r (cid:48) ( t n ) of an actual event to be categorized as a particle of the type 1or 2, the Gatti parameter G is defined as: G = Σ n r (cid:48) ( t n ) w ( t n ) with w ( t n ) = r ( t n ) − r ( t n ) r ( t n ) + r ( t n ) (3.3)and can be used for discrimination. Due to the simplicity and stability of the Gatti analysis, we willuse it as our baseline in this paper.Additionally, also in order to address more subtle problems like e + / e − and e − / γ discrimination,we compare our Gatti results to a neural network (NN) analysis. In principle, the Gatti filter is alinear signal transformation and can be seen as a NN reduced to only input and output layer, withthe time profile replacing the training process. The struture of the implemented NN is shown inFig. 4. Analogous to the Gatti method, the input is a ToF-corrected pulse shape with 400 bins of1 ns in size. Only one hidden dense layer with 20 neurons followed by an activation layer was addedas it turned out that additional hidden layers did not improve the results. The output layer assignsan affiliation probability to each particle type. The cut value was set at equal probabilities of 0.5. Figure 4 . Structure of the neural network applied for the particle identification.
The topology of an event in a large, unsegmented LS detector can be partly recovered from the PMThit information using a method described in [33]. In addition to hit times and respective charges, thetopological reconstruction (TR) requires a knowledge of the reference parameters r ref and t ref , whichdenote one point in space and time, respectively, which the primary particle must have traversed.The parameters r ref and t ref can be obtained e.g. from an independent vertex reconstruction. Thedetection time t hit of a scintillation photon produced at a position r along the particle track andobserved as the k th hit on the j th PMT at position r j can be expressed by t hit = t ref ± | r − r ref | c + | r j − r | v g + t s . (3.4)– 8 –he second term represents the flight time of the particle under the assumption that it moveswith vacuum speed of light c , being subtracted or added depending on the particle reaching r before or after traversing r ref , respectively. The third term considers the time of flight of thescintillation photon, whose group velocity v g depends on its wavelength and the refractive indexof the surrounding medium. Light attenuation as caused e.g. by Rayleigh scattering or photonabsorption is not considered. The non-deterministic contributions from the statistical scintillationprocess and the timing uncertainty of the PMTs are merged in the summand t s , which can thus alsobe negative.Solving Eq. 3.4 for r yields an isochronic surface centered around the PMT at r j . However,the fact that the exact t s is unknown but instead emanates from a probability density function(PDF) of time causes the isochrone to smear out perpendicularly to the surface. The profile of thissmearing is mostly (ignoring dispersion affects during propagation) given by the scintillation timeprofile convoluted with the time response of the PMTs. We always use the scintillation time profileexpected for electrons, although the scintillation time profile depends on the interacting particle. In addition, a filter is applied to this 3D distribution in order to take into account the local probability ε j ( r ) of light to be detected at r j , considering light attenuation and the detector geometry. Theresult, when normalized to 1, is a 3D PDF for the origin of the detected photon, in the followingreferred to as φ j , k ( r ) . Adding up the contributions from all hits and PMTs, i.e. (cid:205) j , k φ j , k ( r ) , yieldsa rough impression of the spatial origin of all detected light. The actual local density Γ em ( r ) ofthe number of emitted photons can be gained from re-weighting (cid:205) j , k φ j , k ( r ) with the inverse of thelocal detection efficiency ε ( r ) . The latter is gained from summing ε j ( r ) over all PMTs, i.e. Γ em ( r ) = (cid:205) j , k φ j , k ( r ) (cid:205) j ε ( r ) . (3.5)The mere superposition of φ j , k ( r ) contributions treats photon emissions as independent inci-dents. In fact all emissions share a common event topology and are thus correlated. This can beutilised by treating the previous result as prior information in further iterations. While re-evaluating φ j , k ( r )| n in the n th iteration, Γ em ( r )| n − is introduced as weighting mask before normalisation,ideally minus the contribution from φ j , k ( r )| n − in order to prevent self enhancement.For high energy O (GeV) events on the one hand, the TR can reveal regions along the particletrack, where an excess of energy deposition has occurred, e.g. due to a hadronic shower. Onthe other hand, for the discussed O (MeV) low energy regime, the TR can, given the O (ns) timeresolution of the PMTs, by no means resolve topological structures on scales below 10 cm.Figure 5 (a) shows a typical example for a low energy TR event in JUNO with the colour coderepresenting a projection of the emission density Γ em ( r ) on the x-y-plane. The units are arbitrarilyscaled. The TR was carried out in 9 iterations for a simulated positron event with 3 MeV kineticenergy. A red cross and a black ring mark the true and reconstructed vertex point, respectively. Twoblack straight lines indicate the simulated tracks of the annihilation gammas. The reconstructedtopology resembles a cloud around the reference point, coming from which the density graduallydecreases. However, the energy depositions from the gammas do not appear as distinct features inthe topology. Instead, the off-centered emissions of scintillation photons cause the cloud to become That is why we expect different reconstruction results for particle with other scintillation time profiles such as alphasand protons. This is where the discrimination power in these cases stems from. – 9 – a) (b)
Figure 5 . Topological reconstruction of a simulated positron event with 3 MeV kinetic energy: (a) projectionof the emission density Γ em ( r ) on the x-y-plane in arbitrary units and (b) its corresponding radial dependencearound the reference point r ref . Details in text. more diffuse and spread a little wider compared to a more point-like electron event. In case ofan alpha or proton event, a similar effect takes place since the increased number of late-photonemissions is associated less closely with the reference point. However, it has to be noted that pulsefeatures in regions of low intensity take effect upon the TR result only marginally. The reason isthat the gradual increase in contrast which is attained during the iteration process goes along withfading of less pronounced topology regions. The TR method in its current state is thus optimisedto expose near-peak variations as anticipated in e + / e − and e − / γ discrimination.The compactness of the reconstructed topology can be studied when building the radial profile f ( r ) as shown in Fig. 5 (b), i.e. plotting the bin content found on average in a radius r around r ref .The gradient defined as g ( r ) = f ( r ) − f ( r + ∆ r ) ∆ r (3.6)over a window with constant size ∆ r takes higher values for more compact topologies. Accordingly,the highest value g max found along r was chosen to be used as a discrimination parameter.Figure 6 shows the direct comparison of g max values for simulated electrons (green) andpositrons (blue). The depicted events were picked at detector radii between 9.5 m and 10.5 m andhave energies between 2.5 MeV and 3.0 MeV. A solid (dashed) black line marks the cut value for arequired efficiency of 90% (50 %) for an electron-like signal. Not only does the positron distributionpeak at a lower value, corresponding to the topology being less point-like, but also does it exhibit ashoulder along its rising edge, caused by the delayed annihilation in ortho-Ps events. Event simulations were carried out with the official Geant4-based JUNO simulation. All analyseswere performed on three distinct datasets: – 10 – igure 6 . Distribution of discrimination parameter g max based on the topological event reconstruction forelectron and positron events at detector radii between 9.5 m and 10.5 m and with visible energies between2.5 MeV and 3.0 MeV. • Dataset 1: the pure MC truth data. A full simulation of the detector was done, implyingthe kinematics during energy deposition, the emission of scintillation light, and the passageof optical photons through the detector media. However, exact knowledge was assumed fordetected hittimes and for the reference point and reference time used in our methods. Thereference point was chosen as the barycentre of energy deposition, the reference time as thetime of first energy deposition. Note that we sometimes refer to this point and time as vertexpoint and vertex time, although it is only identical to the primary vertex in case of point-likeevents and not so for gamma events. This idealised dataset is used in order to explore theabsolute limits of our method• Dataset 2: smearing of vertex and hit times. The consideration of a finite timing resolutionof PMTs is strongly related to the resolution in vertex reconstruction, which furthermoredepends on the number of measured photons, i.e. on visible energy. Based on the eventsfrom Dataset 1, the hit times were smeared with a Gaussian of the width corresponding tothe actual TTS values of the JUNO PMTs. Gaussians were also used to smear the vertexpoint and time. The standard deviations were estimated from the current efforts for vertexreconstruction in JUNO to follow a σ / (cid:112) E / MeV-rule, with E denoting the visible energy:the values of 10 cm and 0.7 ns were chosen for the smearing in each vertex dimension andtime, respectively, both underlying conservative assumptions.• Dataset 3: adding of dark noise. Based on Dataset 2, dark noise is added with a rate of30 kHz for all large PMTs. The direct comparison between Datasets 2 and 3 can reveal theimpact of dark noise on our discrimination methods.For each particle type in the discrimination categories α / β , p / β , e + / e − , and e − / γ , 120k eventswere simulated, 100k of which were taken as training sample and the remaining 20k events forvalidation. The events were spread uniformly over the whole CD with energies according to the– 11 –article α / β e − / p e + / e − e − / γ Energy [MeV] [0.2, 1.5] [0, 2.0] [1.0, 10.0] [0, 3.0]Position uniformly in the whole central detector
Table 2 . Energy range and position distribution of the simulated data samples. intervals quoted in Table 2, selected according to the expected physics applications, as discussed inSec. 2. For e + events, ortho-Ps was considered at a fraction of 54.5% and with a lifetime of 3.08 ns.Since the small PMTs in the CD account for less than 4% of the optical coverage from the largePMTs, it was decided to generally ignore hits on small PMTs in favour of computation time. In order to compare and analyse our methods, the results will be presented based on a fixed scheme.A selection of the figures of merit introduced here will be shown for each event pairing in Sec. 4.A full collection of the plots can be found in the appendix.We define (i) discrimination efficiency (cid:15) sig as the ratio of the number of signal events passing acut and total amount of signal events and (ii) impurity (cid:15) bkg as the ratio of the number of remainingbackground events after the cut and total background events. This implies that neither (cid:15) sig nor (cid:15) bkg depend on the actual ratio of signal to background events.Impurity will be plotted over efficiency in a fixed energy and radius range. Three plots,representing our discrimination methods, will be shown, each containing three curves for theanalysed datasets.Efficiency and impurity will both be plotted as a function of energy at a fixed level of (cid:15) bkg and (cid:15) sig , respectively. Note that this configuration allows to draw equivalent conclusions for switchingsignal and background by simply switching the labels efficiency and impurity and reversing boththeir axes. All datasets will be presented in order to analyse the differences between ideal andrealistic data. Note that for the TR method also the radius range in the CD was fixed to be ( ± . ) m, since the cut parameter was found to change with detector radius R . The cut value forthe NN and Gatti analyses were determined over the whole detector.Efficiency and impurity will also be plotted over the CD volume, parametrised by R , ina defined energy range. An additional horizontal axis indicates the corresponding R -values fororientation. The discussion within Sec. 4 is limited to the most realistic Dataset 3. α / β and p / β Discrimination
Alphas and protons, although showing individual quenching behaviour due to different charges,cause scintillation light to be emitted very similarly over time, which is reflected by almost identicaltime constants and weights in Table 1. Accordingly, no differences are expected when comparingthe results for discrimination against electrons on the basis of coinciding visible energy. Our resultsare indeed congruent within the tested energy ranges. Here, we show the p / β results which cover awider energy range and point out that our conclusions equally apply to the respective plots for α / β appended to this paper (see Appendix A). – 12 – a) Gatti (b) NN(c) TR Figure 7 . Impurity as a function of efficiency for α / β and p / β discrimination. The results were obtainedfor visible energies between 1.5 MeV and 2.0 MeV. A discrimination between α and p on the one hand and β on the other can be considered astraightforward task due to the clear distinction features in the time profiles (Fig. 3 (a)) and alsobased on the experience in other experiments, as previously discussed. In our study, electron eventswere in either case treated as signal. The obtained level of background impurity was plotted as afunction of signal efficiency in Fig. 7 with the Gatti (a), NN (b), and TR (c) method. The events havevisible energies between 1.5 MeV and 2.0 MeV. Each plot contains three curves representing thedifferent datasets. The Gatti and NN method have no apparent difficulty in event classification. Inthe NN case, only the data point at very high efficiency above 95% registers a non-zero backgroundcontamination for Dataset 3. The TR parameter, which was designed and optimised for e + / e − discrimination, is also sensible to p / β discrimination, however, it performs weaker than the directmethods. This is related to a known feature during the TR iteration process, which is the tendencyof intense topology regions to attract the probability contributions that would technically correlatebest with less pronounced regions. Since the most striking differences in pulse shape appear at latetimes, where the pulse is low, the TR shows only weak sensitivity here. Vanishing impurities below50% efficiency and a steep rise at high efficiencies show that the TR parameter is usable for pickingpure signal samples but, other than NN and Gatti, inappropriate for highly efficient background– 13 – c) Dataset 1 (d) Dataset 2(g) Dataset 3 (h) Dataset 3 Figure 8 . Performance of the α / β and p / β discrimination from all three methods. Impurity was obtained atefficiency fixed to 90% while efficiency was obtained at impurity fixed to 10 %. (a), (b), and (c) show resultswith Datasets 1, 2, and 3, respectively, as a function of visible energy. (d) shows the performance dependingon the detector radius. – 14 –uts. The deterioration is strongly being amplified by including TTS and vertex (Dataset 2) and theaddition of dark noise (Dataset3).A direct comparison between all three methods is demonstrated in Fig. 8. Panels (a), (b),and (c) show impurity and efficiency as a function of energy for Datasets 1, 2, and 3, respectively,while panel (d) shows the radius dependence for Dataset 3. Efficiency was determined at a fixedimpurity level of 10%. Impurity was determined with the required efficiency set to 90%. It can beobserved as a general trend that higher energies, which imply an increase in p.e. statistics, favour theprediction power in PID. Even in the ideal Dataset 1 a clean data cut is achieved only above 1 MeV.Impurities below this value rise fast towards lower energies, while a fixed level of impurity wouldgo along with an according drop in efficiency. The transitions from Datasets 1 to 3 shift this edge tohigher energies. In direct comparison, the NN results mildly exceed those achieved with the Gattianalysis. The TR method is suffering more than the direct methods from the lack of p.e. statisticsat low energies. In Dataset 3, which represents the most realistic data, the TR method loses itsprediction power below 0.6 MeV. Here, the contribution of ∼
200 dark hits within the critical 400 nsof pulse shape weighs heavy compared to the ∼ and 3500 m . A look atthe p.e. yield over detector radius displayed in Fig. 2 reveals that in fact least light is expected fromthe innermost and outermost detector regions, meaning less statistics for the analysis and having asimilar effect as observed at lower energies. In contrast to the direct methods, the TR approach losesall prediction power above R ≈ , corresponding to R ≈ . e + / e − Discrimination
This analysis regards electron events as signal and positron events as background. However,depending on the physics case the requirement can be vice versa.The behaviour of the e + / e − discrimination differs substantially from the previous results,which build on features in the light emission curves rather than on characteristic topologies. Thedifferences between e + and e − pulse shapes are much less pronounced as can be spotted in Fig. 3.Accordingly, the depicted impurities in Fig. 9 which were obtained for events with visible energiesbetween 2.75 MeV and 3.25 MeV are on a higher level than for α / β and p / β . Especially the Gattianalysis turns out to be inappropriate for the task: the slight sensitivity observable in Dataset 1is almost lost when going to realistic data, expressed by near-linear curves with a slope close to1. However, the similar results for the NN and TR method prove that a considerable amount ofbackground can be removed by the cost of much less signal, e.g. around 30% impurity at 90%efficiency in both methods for Dataset 3. Impurity rises faster towards high efficiencies. Thus, forphysics studies with high sample rates at hand, it could pay off to lower the efficiency requirementsin favour of an increased signal-to-noise ratio. A large gap appears between the points for Datasets– 15 – a) Gatti (b) NN(c) TR Figure 9 . Impurity as a function of efficiency for e + / e − discrimination. The results were obtained for visibleenergies between 2.75 MeV and 3.25 MeV. e + kinetic energy causing the central ionisation, therelative weight of the off-center energy deposition of the two annihilation gammas of 1.022 MeVtotal energy, decreases. As soon as vertex and TTS smearing as well as dark noise enter the data(Datasets 2 and 3, respectively), the NN performance becomes worse also towards the low end ofthe energy spectrum. Apparently, the deterioration of data quality cannot fully be compensatedfor by statistics at these energies. As a result, the most sensitive region lies around 3 MeV. Thisactually meets the experimental focus for solar B neutrinos which lies between 2 MeV and 5 MeV.In absolute numbers, the impurities obtained for Dataset 3 do not fall below 5% at 50% efficiency,which rules out an event-by-event discrimination. TR and NN produce very similar results between2 MeV and 3 MeV. With rising energy, the TR values depart further from the NN values.– 16 – c) Dataset 1 (d) Dataset 2(g) Dataset 3 (h) Dataset 3
Figure 10 . Performance of the e + / e − discrimination from all three methods. Impurity was obtained atefficiency fixed to 50%, while efficiency was obtained at impurity fixed to 20 %. (a), (b), and (c) show resultswith Datasets 1, 2, and 3, respectively, as a function of visible energy. (d) shows the performance dependingon the detector radius. – 17 –oncerning the impact of dark noise, it can be concluded that the effect is less serious thanfound for the preceding event categories. This can be traced back to fact that here the differencesmainly are encrypted in the early part of the hit spectrum where the pulse peaks. The same reasonexplains why the TR method shows overall better results and is compatible with the NN. The Gattimethod on the other hand turns out to be unsuitable for e + / e − discrimination.The radius dependence shown in Fig. 10(d) is consistent with the previously discussed eventcategories. e − / γ Discrimination (a) Gatti (b) NN(c) TR
Figure 11 . Impurity as a function of efficiency for e − / γ discrimination. The results were obtained for visibleenergies between 2.0 MeV and 2.5 MeV. The discrimination of e − signal against γ background was expected to be the most challengingof the investigated categories. The efficiency scan is shown in Fig. 11 for visible energies between2.0 MeV and 2.5 MeV. Like for e + / e − , the NN and TR method prove to be sensitive to the task.Again, a large gap between Datasets 1 and 2 indicates that in future detectors much potential canstill be exploited by more accurate light sensors. The Gatti parameter, on the other hand, is hardlyable to discriminate at all. – 18 – c) Dataset 1 (d) Dataset 2(g) Dataset 3 (h) Dataset 3 Figure 12 . Performance of the e − / γ discrimination from all three methods. Impurity was obtained atefficiency fixed to 50%, while efficiency was obtained at impurity fixed to 20 %. (a), (b), and (c) show resultswith Datasets 1, 2, and 3, respectively, as a function of visible energy. (d) shows the performance dependingon the detector radius. – 19 –nergy and radius dependence were studied at 50% efficiency and 10% impurity, the resultsare shown in Fig. 12. The data situation is similar to the e + / e − case because the critical featuresin the time profiles arise at early times. Within the energy range shared by the e + / e − and e − / γ study, i.e. between 1 MeV and 3 MeV, the TR parameter reaches comparable results. Below 1 MeV,low statistics deteriorate the performance. The NN is less powerful in e − / γ but approaches the TRresults with rising energy. Since gammas spread their energy over a wider region with rising energy,a continuous decrease in impurity is expected even beyond the investigated energy region. Thisactually does apply for certain gammas not descending from the natural decay chains, e.g. 6 MeVand 8.5 MeV gammas from neutron captures in the stainless steel surrounding JUNO’s acrylic CDsphere. The potential for event discrimination in JUNO was extensively studied on the basis of three distinctmethods and different event pairings. Our studies concerning α / β and p / β promise very cleanand reliable cuts even at very low energies around 0.2 MeV. A high p.e. yield brings also e + / e − discrimination into reach, although higher impurity rates only allow for statistical classificationhere. We point out that better results can be achieved for cosmogenic C background, whose β + -decay is followed with 1 ns delay by a 718 keV γ transition [34]. A separation between e − and γ events was considered more challenging but was actually found to be feasible for energies above1 MeV. It needs to be investigated in a dedicated study how our discrimination would influenceJUNO’s sensitivities in the solar neutrino sector. While the gamma contamination is expected togrow exponentially with detector radius, the usable volume is linked with its third power, meaningthat already small expansions in fiducial radius would lead to a massive gain in the amount of data.However, the accuracy is not high enough to expand JUNO’s fiducial volume significantly in solarneutrino studies.Except for e + / e − , all event pairings showed a continuous trend to gain in discriminationperformance with visible energy. The former case however differentiates from the others since thedecisive γ component from e + annihilation is constant in terms of energy deposition and recedesbehind the contribution from kinetic energy. The examination of radius dependence revealed acorrelation between cut performance and number of detected photons. This is in accordance withthe observed energy dependence and causes the best results to show up at detector radii between10 m and 16 m.Apart from the fraction of direct light and the absolute p.e. yield, also the technical equipmentinfluences data quality. We found that timing uncertainties of PMTs, in turn being related to thevertex resolution, have an impact on the discrimination. Dark noise affects the results particularlyat low energies, where they significantly reduce the relative fraction of direct photon signals. Thoseparameters need to be considered in the design for future detectors like THEIA [2], where photosensors with ∼
100 ps resolution represent a design option. The potential lying in an optimiseddetector can be learned from the big gap which still exists between our results with the ideal Dataset1 and the more realistic Datsets 2 and 3.A direct comparison between the discrimination methods shows the power of the applied NNin spite of its simple architecture, as it proves to be sensitive to all studied cases. The Gatti and TR– 20 –ethod played out their strengths in different disciplines. Gatti returns good results in α / β and e − / p discrimination, both relying on characteristics in the time spectrum of scintillation which show upespecially in the tail region of the time profiles. The TR performs to its full potential in e + / e − and e − / γ discrimination, where the distinction features manifest themselves around the profile peak.Further efforts in the development of the TR need to focus on the performance towards the detectoredge, where distortions momentarily impedes PID. Acknowledgements
This work was supported by the Deutsche Forschungsgemeinschaft and the Helmholtz Association.Furthermore, we would like to thank the JUNO collaboration for providing the Monte Carlo softwarewith which we carried out the simulations for this study.– 21 –
Full collection of plots for α / β discrimination (a) Gatti (b) NN (c) TR Figure 13 . Impurity as a function of efficiency for α / β discrimination. The results were obtained for visibleenergies between 1.0 MeV and 1.5 MeV. (d) Dataset 1 (e) Dataset 2 (f) Dataset 3 Figure 14 . Energy dependence of the α / β discrimination from all three methods. Impurity was obtained atefficiency fixed to 90%, while efficiency was obtained at impurity fixed to 10 %. – 22 – d) Dataset 1 (e) Dataset 2 (f) Dataset 3 Figure 15 . Radius dependence of the α / β discrimination from all three methods. Impurity was obtained atefficiency fixed to 90%, while efficiency was obtained at impurity fixed to 10 %. – 23 – Full collection of plots for p / β discrimination (a) Gatti (b) NN (c) TR Figure 16 . Impurity as a function of efficiency for p / β discrimination. The results were obtained for visibleenergies between 1.5 MeV and 2.0 MeV. (d) Dataset 1 (e) Dataset 2 (f) Dataset 3 Figure 17 . Energy dependence of the p / β discrimination from all three methods. Impurity was obtained atefficiency fixed to 90%, while efficiency was obtained at impurity fixed to 10 %. – 24 – d) Dataset 1 (e) Dataset 2 (f) Dataset 3 Figure 18 . Radius dependence of the p / β discrimination from all three methods. Impurity was obtained atefficiency fixed to 90%, while efficiency was obtained at impurity fixed to 10 %. – 25 – Full collection of plots for e + / e − discrimination (a) Gatti (b) NN (c) TR Figure 19 . Impurity as a function of efficiency for e + / e − discrimination. The results were obtained forvisible energies between 2.75 MeV and 3.25 MeV. (d) Dataset 1 (e) Dataset 2 (f) Dataset 3 Figure 20 . Energy dependence of the e + / e − discrimination from all three methods. Impurity was obtainedat efficiency fixed to 50%, while efficiency was obtained at impurity fixed to 20 %. – 26 – d) Dataset 1 (e) Dataset 2 (f) Dataset 3 Figure 21 . Radius dependence of the e + / e − discrimination from all three methods. Impurity was obtainedat efficiency fixed to 50%, while efficiency was obtained at impurity fixed to 20 %. – 27 – Full collection of plots for e − / γ discrimination (a) Gatti (b) NN (c) TR Figure 22 . Impurity as a function of efficiency for e − / γ discrimination. The results were obtained for visibleenergies between 2.0 MeV and 2.5 MeV. (d) Dataset 1 (e) Dataset 2 (f) Dataset 3 Figure 23 . Energy dependence of the e − / γ discrimination from all three methods. Impurity was obtained atefficiency fixed to 50%, while efficiency was obtained at impurity fixed to 20 %. – 28 – d) Dataset 1 (e) Dataset 2 (f) Dataset 3 Figure 24 . Radius dependence of the e − / γ discrimination from all three methods. Impurity was obtained atefficiency fixed to 50%, while efficiency was obtained at impurity fixed to 20 %. References [1] Fengpeng An et al. Neutrino Physics with JUNO.
J. Phys. G , 43(3):030401, 2016.[2] M. Askins et al. Theia: An advanced optical neutrino detector, 2019.[3] John F. Beacomand et al. Physics prospects of the jinping neutrino experiment.
Chinese Physics C ,41(2):023002, feb 2017.[4] Christine Kraus and Simon J.M. Peeters. The rich neutrino programme of the sno+ experiment.
Progress in Particle and Nuclear Physics , 64(2):273 – 277, 2010. Neutrinos in Cosmology, in Astro,Particle and Nuclear Physics.[5] A. Gando and et al. Search for majorana neutrinos near the inverted mass hierarchy region withkamland-zen.
Phys. Rev. Lett. , 117:082503, Aug 2016.[6] H.O. et al. Back. Pulse-shape discrimination with the counting test facility.
Nuclear Instruments andMethods in Physics Research Section A: Accelerators, Spectrometers, Detectors and AssociatedEquipment , 584(1):98âĂŞ113, Jan 2008.[7] Borexino Collaboration. Final results of borexino phase-i on low-energy solar neutrino spectroscopy.
Phys. Rev. D , 89:112007, Jun 2014.[8] Y. Abe and et al. Ortho-positronium observation in the double chooz experiment.
Journal of HighEnergy Physics , 2014(10), Oct 2014.[9] Stefan Wagner.
Energy non-linearity studies and pulse shape analysis of liquid scintillator signals inthe Double Chooz experiment . PhD thesis, Heidelberg U., 2014.[10] R. Möllenberg, F. von Feilitzsch, D. Hellgartner, L. Oberauer, M. Tippmann, V. Zimmer, J. Winter,and M. Wurm. Detecting the diffuse supernova neutrino background with lena.
Phys. Rev. D ,91:032005, Feb 2015.[11] S. Fukuda, Y. Fukuda, and et al. The super-kamiokande detector.
Nuclear Instruments and Methodsin Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment ,501(2):418 – 462, 2003. – 29 –
12] S. Abe, S. Enomoto, K. Furuno, and et al. Production of radioactive isotopes through cosmic muonspallation in kamland.
Phys. Rev. C , 81:025807, Feb 2010.[13] G Bellini and et al. Cosmogenic backgrounds in borexino at 3800 m water-equivalent depth.
Journalof Cosmology and Astroparticle Physics , 2013(08):049–049, aug 2013.[14] F. P. An, A. B. Balantekin, and et al. Measurement of electron antineutrino oscillation based on 1230days of operation of the daya bay experiment.
Phys. Rev. D , 95:072006, Apr 2017.[15] Abrahão T. Almazan H. de Kerret, H. et al. Yields and production rates of cosmogenic Li and Hemeasured with the Double Chooz near and far detectors.
Journal of High Energy Physics , 53(11), Jul2018.[16] M. Agostini et al. Comprehensive measurement of pp -chain solar neutrinos. Nature ,562(7728):505–510, 2018.[17] K. Abe et al. Solar Neutrino Measurements in Super-Kamiokande-IV.
Phys. Rev. D , 94(5):052010,2016.[18] L. Wolfenstein. Neutrino Oscillations in Matter.
Phys. Rev. D , 17:2369–2374, 1978.[19] S.P. Mikheyev and A.Yu. Smirnov. Resonance Amplification of Oscillations in Matter andSpectroscopy of Solar Neutrinos.
Sov. J. Nucl. Phys. , 42:913–917, 1985.[20] The Collaboration, Matteo Agostini, K. Altenmueller, S. Appel, Viktor Atroshchenko, ZaraBagdasarian, David Basilico, G.B. Bellini, Jay Benziger, D. Bick, G. Bonfini, Diane Bravo,B. Caccianiga, F. Calaprice, A. Caminata, Silvia Caprioli, Marco Carlini, P. Cavalcante,A. Chepurnov, and G. Zuzel. Improved measurement of 8b solar neutrinos with 1.5 kt y of borexinoexposure. 09 2017.[21] Itaru Shimizu. Kamland: geo-neutrino measurement in japan.
ANNALS OF GEOPHYSICS , 60(1),2017.[22] Jie Zhao, Ze-Yuan Yu, Jiang-Lai Liu, Xiao-Bo Li, Fei-Hong Zhang, and Dong-Mei Xia. 13c( α ,n)16obackground in a liquid scintillator based neutrino experiment. Chinese Physics C , 38(11):116201, nov2014.[23] J. Cheng J. Sawatzki. Detecting the diffuse supernova neutrino background in juno. 2020.[24] R. MÃűllenberg, F. von Feilitzsch, D. Hellgartner, L. Oberauer, M. Tippmann, V. Zimmer, J. Winter,and M. Wurm. Detecting the diffuse supernova neutrino background with lena.
Physical Review D ,91(3), Feb 2015.[25] C. A. Parker and C. G. Hatchard. Triplet-singlet emission in fluid solutions. phosphorescence ofeosin.
Trans. Faraday Soc. , 57:1894–1904, 1961.[26] C. A. Parker and C. G. Hatchard. Delayed fluorescence from solutions of anthracene andphenanthrene.
Proceedings of the Royal Society of London. Series A, Mathematical and PhysicalSciences , 269(1339):574–584, 1962.[27] G. Laustriat. The luminescence decay of organic scintillators.
Molecular Crystals , 4(1-4):127–145,1968.[28] J. B. Birks.
The theory and practice of scintillation counting / by J.B. Birks . Pergamon Press Oxford[England], 1964.[29] D. Franco, G. Consolati, and D. Trezzi. Positronium signature in organic liquid scintillators forneutrino experiments.
Phys. Rev. C , 83:015504, Jan 2011. – 30 –
30] G. Consolati, D. Franco, S. Hans, C. Jollet, A. Meregaglia, S. Perasso, A. Tonazzo, and M. Yeh.Characterization of positronium properties in doped liquid scintillators.
Phys. Rev. C , 88:065502, Dec2013.[31] Mario Schwarz, Sabrina M. Franke, Lothar Oberauer, Miriam D. Plein, Hans Th.J. Steiger, and MarcTippmann. Measurements of the lifetime of orthopositronium in the lab-based liquid scintillator ofjuno.
Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers,Detectors and Associated Equipment , 922:64 – 70, 2019.[32] E. Gatti and F. De Martini.
A new linear method of discrimination between elementary particles inscintillation counters . International Atomic Energy Agency (IAEA): Brueder Rosenbaum, 1962.[33] B.S. Wonsak, C.I. Hagner, D.A. Hellgartner, K. Loo, S. Lorenz, D.J. MeyhÃűfer, L. Oberauer,H. Rebber, W.H. Trzaska, and M. Wurm. Topological track reconstruction in unsegmented,large-volume liquid scintillator detectors.
Journal of Instrumentation , 13(07):P07005–P07005, jul2018.[34] Henning Rebber.
Event Discrimination with Topological 3D Reconstruction at MeV Energies in theJUNO Experiment . PhD thesis, University of Hamburg, 11 2019.. PhD thesis, University of Hamburg, 11 2019.