Identification of low energy nuclear recoils in a gas TPC with optical readout
Elisabetta Baracchini, Luigi Benussi, Stefano Bianco, Cesidio Capoccia, Michele Arturo Caponero, Gianluca Cavoto, Andre Cortez, Igor Abritta Costa, Emanuele Di Marco, Giulia D'Imperio, Giorgio Dho, Fabrizio Iacoangeli, Giovanni Maccarrone, Michela Marafini, Giovanni Mazzitelli, Andrea Messina, Rafael Antunes Nobrega, Aldo Orlandi, Emiliano Paoletti, Luciano Passamonti, Fabrizio Petrucci, Davide Piccolo, Daniele Pierluigi, Davide Pinci, Francesco Renga, Filippo Rosatelli, Alessandro Russo, Giovanna Saviano, Roberto Tesauro, Sandro Tomassini
IIdentification of low energy nuclear recoils in a gasTPC with optical readout
E Baracchini , , L Benussi , S Bianco , C Capoccia , MCaponero , , G Cavoto , , A Cortez , , I A. Costa , E DiMarco , G D’Imperio , G Dho , , F Iacoangeli , G Maccarrone ,M Marafini , , G Mazzitelli , A Messina , , R A. Nobrega , AOrlandi , E Paoletti , L Passamonti , F Petrucci , , D Piccolo ,D Pierluigi , D Pinci , F Renga , F Rosatelli , A Russo , GSaviano , R Tesauro , and S Tomassini Gran Sasso Science Institute, L’Aquila, Italy INFN, Laboratori Nazionali del Gran Sasso, Assergi, Italy INFN, Laboratori Nazionali di Frascati, Frascati, Italy ENEA Centro Ricerche Frascati, Frascati, Italy INFN, Sezione di Roma, Roma, Italy Dipartimento di Fisica, Sapienza Università di Roma, Roma, Italy Universidade Federal de Juiz de Fora, Juiz de Fora, Brasil Museo Storico della Fisica e Centro Studi e Ricerche “Enrico Fermi”, Roma, Italy Dipartimento di Matematica e Fisica, Università Roma TRE, Roma, Italy Dipartimento di Ingegneria Chimica, Materiali e Ambiente, Sapienza Università diRoma, Roma, ItalyE-mail: [email protected]
24 July 2020
Abstract.
The search for a novel technology able to detect and reconstruct nuclearrecoil events in the keV energy range has become more and more important as longas vast regions of high mass WIMP-like Dark Matter candidate have been excluded.Gaseous Time Projection Chambers (TPC) with optical readout are very promisingcandidate combining the complete event information provided by the TPC techniqueto the high sensitivity and granularity of last generation scientific light sensors. A TPCwith an amplification at the anode obtained with Gas Electron Multipliers (GEM) wastested at the Laboratori Nazionali di Frascati. Photons and neutrons from radioactivesources were employed to induce recoiling nuclei and electrons with kinetic energy inthe range [1–100] keV. A He-CF (60/40) gas mixture was used at atmospheric pressureand the light produced during the multiplication in the GEM channels was acquired bya high position resolution and low noise scientific CMOS camera and a photomultiplier.A multi-stage pattern recognition algorithm based on an advanced clustering techniqueis presented here. A number of cluster shape observables are used to identify nuclearrecoils induced by neutrons originated from a AmBe source against X-ray Fe photo-electrons. An efficiency of 18% to detect nuclear recoils with an energy of about 6keV is reached obtaining at the same time a 96% Fe photo-electrons suppression.This makes this optically readout gas TPC a very promising candidate for futureinvestigations of ultra-rare events as directional direct Dark Matter searches. a r X i v : . [ phy s i c s . i n s - d e t ] J u l
1. Introduction
The advent of a market of high position resolution and single photon light sensors canopen new opportunity to investigate ultra-low rate phenomena as Dark Matter (DM)particle scattering on nuclei in a gaseous target.The nature of DM is still one of the key issues to understand our Universe [1, 2].Different models predict the existence of neutral particles with a mass of few GeVs orhigher that would fill our Galaxy [3–6]. They could interact with the nuclei present inordinary matter producing highly ionizing nuclear recoils but with a kinetic energy assmall as few keVs. Moreover, given the motion of the Sun in the Milky Way towards theCygnus constellation, such nuclear recoils would exhibit in galactic coordinate a dipoleangular distribution in a terrestrial detector [7]. In this paper the use of a scientificCMOS camera to capture the light emitted by Gas Electron Multipliers (GEMs) in aTime Projection Chamber (TPC) device is described. The GEMs are located in theTPC gas volume at the anode position and are used to amplify the ionization producedin the gas by the nuclear recoils and other particles. A secondary scintillation light isproduced in the amplification of the avalanche process by the GEM. This light and itsspatial distribution is reconstructed in the detector and characterized by means of aclustering algorithm.Different type of particles will produce distinctive and diverse patterns of ionizationcharge, and therefore of light emitted by the GEMs, given the different way they depositenergy and interact with matter. Therefore, nuclear recoils can be efficiently identifiedand separated from different kinds of background down to a few keV kinetic energy. Thestudy of the optical readout of a TPC has been recently conducted with several smallsize prototypes (NITEC [8], ORANGE [9, 10],
Lemon [11–13]) with various particlesources, in the context of the
Cygno project [14, 15]. In the following, the study ofnuclear recoils excited by neutrons from an AmBe source and electron recoils from a Fe source in the gas volume of the
Lemon prototype is presented.
2. Experimental layout
A 7 liter active sensitive volume TPC (named
Lemon ) was employed to detect theparticle recoils. A sketch (not to scale) of the detector setup is shown in Fig. 1 (left),while an image of the detector in the experimental area is shown in Fig. 1 (right). Thesensitive volume where the ionization electrons are drifting features a 200 ×
240 mm elliptical field cage with a 200 mm distance between the anode and the cathode. Theanode side is instrumented with a × mm rectangular triple GEM structure.Standard LHCb-like GEMs (70 µ m diameter holes and 140 µ m pitch) [16] were usedwith two 2 mm wide transfer gaps between them. The light emitted from the GEMs isdetected with an ORCA-Flash 4.0 camera [17] through a × × mm transparentwindow and a bellow of adjustable length. This camera is positioned at a 52 cm distancefrom the outermost GEM layer and is based on a sCMOS sensor with high granularity Figure 1.
Left: the
Lemon prototype with its 7 liter sensitive volume (A), the PMT(B), the adjustable bellow (C) and the sCMOS camera with its lens (D). Right:
Lemon with the lead shield around the drift volume cage. The sCMOS camera (on the front)is looking at the GEMs through a blackened bellow. ( × pixels), very low noise (around two photons per pixel), high sensitivity (70%quantum efficiency at 600 nm) and good linearity [18]. This camera is instrumentedwith a Schneider lens, characterized by an aperture f/0.95 and a focal length of 25 mm.The lens is placed at a distance d = 50 . cm from the last GEM in order to obtain ade-magnification ∆ = ( d/f ) − . to image a surface . × . cm onto the . × . cm sensor. In this configuration, each pixel is therefore imaging an effectivearea of × µ m of the GEM layer. The fraction of the light collected by thelens is evaluated to be . × − [18]. A semi-transparent mesh was used as a cathodein order to collect light on that side also with a × mm HZC Photonics XP3392 photomultiplier [19] (PMT) detecting light through a transparent × × mm fusedsilica window. More details on the Lemon detector can be found in Ref. [20].A 5 cm thick lead shielding was mounted around the
Lemon field cage to reducethe environmental natural radioactivity background. From the measurements of theGEM current with and without the lead shielding, a factor two reduction in the totalionization within the sensitive volume, very likely due to environmental radioactivity,was estimated.
3. Particle images in the
Lemon gas volume
The
Lemon detector was operated in an overground location at Laboratori Nazionalidi Frascati (LNF) with a He-CF (60/40) gas mixture at atmospheric pressure, thetriple GEM system set at a voltage across each GEM sides of 460 V and a transfer fieldbetween the GEM layers of 2.5 kV/cm. A six-independent-HV-channels CAEN A1257 module ensured stability and monitored the bias currents with a precision of 20 nA.The gas mixture was kept under continuous flow of about 200 cc/min and with theGEMs operated at a . × electric gain. The typical photon yield for this type of gasmixtures has been measured to be around 0.07 photons per avalanche electron [18,21,22]and therefore the overall light gain is about . The field cage was powered by a CAEN N1570 [23], generating an electric field of 0.5 kV/cm.The motion of particles within the gas mixtures was studied by means of differentsimulation tools. In particular,
Garfield [24, 25] program was used to evaluate thetransport properties for ionization electrons in the sensitive volume for an electric fieldof 500 V/cm.Given the diffusion in the gas, ionization electrons produced at a distance z fromthe GEM will distribute over a region on the GEM surface, having a Gaussian transverseprofile with a σ given by: σ = (cid:113) σ ⊕ D · z, (1)where D is the transverse diffusion coefficient, whose value at room temperature140 µ m / √ cm was obtained with a simulation. The value of σ was measured to beabout 300 µ m [26, 27]. Therefore, in average, a point-like ionization will result in a spotof [3–4] mm .The expected effective ranges of electron and nuclear recoils were evaluatedrespectively with Geant4 [28] and with
Srim [29] simulation programs. The recoilrange estimated from simulation, as a function of the impinging particle kinetic energy,is shown in Fig. 2 for electrons and -as an example- for He-nuclei. These results show m ] m D R ange [ He recoile recoil
Figure 2.
Average ranges for electron and He-nucleus recoils as a function of theirinitial kinetic energy. that: • He-nuclei recoils have a sub-millimeter range up to energies of 100 keV and are thusexpected to produce bright spots with sizes mainly dominated by diffusion; • low energy (less than 10 keV) electron recoils are in general longer than He-nucleusrecoils with same energy and are expected to produce sparse and less intense spot-like signals. For a kinetic energy of 10 keV, the electron range becomes longer than1 mm and for few tens of keV, tracks of few cm are expected.The images collected by the sCMOS camera contain several instances of theparticles tracks described above. The sCMOS sensor was operated in continuous modewith a global exposure time of 30 ms. Example images are shown in Fig. 3. x (pixels) y ( p i x e l s ) Image after zero suppression x (pixels) y ( p i x e l s ) Image after zero suppression
Figure 3.
Two example pictures taken with the sCMOS camera with a 30 ms exposuretime and with a common noise level subtracted (zero suppression), belonging to a datataking run without any artificial source. Left: cosmic tracks and natural radioactivitysignals are present. Right: only two long cosmic rays tracks are visible. The coordinatesare defined such that the vertical direction is along the y-axis and cosmic rays areexpected to come from the top of the figure.
The PMT waveform was sent to a digitizer board with a sampling frequency of4 GS/s. The trigger scheme of the detector is based on the PMT signal: if, during theexposure time window, the PMT waveform exhibits a peak exceeding a threshold of80 mV, it is acquired in a time window of 25 µ s and the corresponding sCMOS imageis stored. The digitizer is operated in single-event mode. No more than one 25 µ s longPMT waveform is recorded in each sCMOS exposure time, even if during the sCMOSexposure time several PMT signals are produced. Therefore, the PMT information wasmainly exploited only to select events with a cosmic ray track.Several light spots are visible with different ionization patterns due to different typesof particles interacting in the gas. Figure 3 (left) shows an image with typical long tracksfrom cosmic rays traveling through the full gas volume, where clusters of light with largerenergy deposition are clearly visible, superimposed to low energy electrons, very likelydue to natural radioactivity. Figure 3 (right) shows an example of a cleaner event withtwo straight cosmic ray tracks, that can be used for energy calibration purposes, sincethe energy releases along the path, dE/dx , can be predicted given the gas mixture andpressure.Two different artificial radioactive sources were employed for testing and studyingthe detector responses.A neutron source, based on a . × MBq activity
Am source contained in aBeryllium capsule (AmBe) was placed at a distance of 50 cm from the sensitive volumeside. Because of the interactions between α particles produced by the Am and theBeryllium nuclei, the AmBe source isotropically emits: • photons with an energy of 59 keV produced by Am; • neutrons with a kinetic energy mainly in a range [1–10] MeV • photons with an energy of 4.4 MeV produced along with neutrons in the interactionbetween α and Be nucleus.The presence of a lead shield around the sensitive volume absorbed almost completelythe 59 keV photon component. A small faction of it reached the gas through small gapsaccidentally present between the lead bricks.A Fe source emitting
X-rays with a main energy peak at 5.9 keV. This is thestandard candle for calibration and performance evaluation of
Lemon , and its extensiveuse is documented in Ref. [30].Four different sets of runs have been recorded: (i) without any source and no electricsignal amplification in the GEMs, to study the sensor electronic noise; (ii) without anysource, but the detector fully active, to study the ambient background, mainly muonsfrom cosmic rays and natural radioactivity; runs with either the (iii) Fe source, tostudy the detector response to a known signal or with the AmBe source, to study the
Lemon performances in presence of nuclear recoils.Figure 4 shows images recorded with the same 30 ms exposure time, in presenceof one of the two sources. The left panel shows an example of several light spots,characteristic of energy deposits due to Fe low energy photons. The right panel showsa frame recorded in presence of the AmBe radioactive source: the short and bright trackwell visible in the center is very likely due to an energetic nuclear recoil induced by aneutron scattering. x (pixels) y ( p i x e l s ) Image after zero suppression x (pixels) y ( p i x e l s ) Image after zero suppression
Figure 4.
Two pictures taken with the sCMOS camera with a 30 ms exposure time.Left: picture taken in presence of Fe radioactive source. Right: a nuclear recoilcandidate is present, in an image with AmBe radioactive source, together with signalsfrom natural radioactivity. The coordinates are defined such that the vertical directionis along the y-axis.
4. Cluster reconstruction algorithm
The light produced in the multiplication process through the GEMs and detected bythe sCMOS sensor is associated in clusters of neighboring pixels. This is achieved byfollowing the trail of energy deposition of the particle traveling through the gas of thesensitive volume. The energy released as ionization electrons is estimated by the amountof the light collected by the sensor. In the range from few keV to tens of keV for stoppedelectrons, the deposited energy is equivalent to their total kinetic energy, while forstopped nuclei it represent only a fraction of their initial kinetic energy. Therefore, it isof primary importance to have a reconstruction algorithm that includes all the camerapixels hit by the real photons originating from the energy deposits, while rejecting mostof the electronic noise from the camera sensor. Noise can either create fake clustersor, more likely, add pixels in the periphery of clusters originated by real photons, thusbiasing the energy estimate. Possible additional noise, arising for example from GEMstages, was already demonstrated be negligible [30].The energy reconstruction follows a three-steps procedure: the single-pixel noisesuppression is briefly described in Section 4.1. This is followed by the proper clustering:first the algorithm to form basic clusters from single small deposits is described inSection 4.2, then the supercluster method, aiming to follow the full particle track, andseeded by the basic clusters found in the previous step, is described in Section 4.3.The results of this paper are based on the properties of the reconstructedsuperclusters and are described in Section 5.
The electronic noise of the sensor was estimated in data-taking runs acquired with thesensor in complete dark, obtained by covering the camera lens with its own cap or,equivalently, lowering the voltage across the GEM electrodes to 300 V ( pedestal runs).The latter option, which was demostrated to be fully equivalent to the former, is avaluable method to measure the sensor noise periodically, and track its evolution, duringthe periods without data taking of the
Cygno experiment.For each pixel, the pedestal was computed as the average of the counts over manyframes, while the electronic noise was estimated as their standard deviation (SD). Thedistribution of the pixels SD is shown in Fig. 5. The mode of this distribution is about1.8 photons per pixel, but a tail is present, with pixels having a noise of more than5 photons per pixel. For such pixels, a very non-Gaussian distribution was observed,while for the pixels in the bulk of the distribution, the pedestal distribution followeda Gaussian shape. To form the pedestal-subtracted image, the pedestal mean µ i wassubtracted to the image for each i th pixel, to account for the non-uniformity of thepedestal mean across the sensor. An initial noise suppression was applied by neglectingthe pixels with counts less than . SD i . On such pedestal-subtracted zero-suppressedimages an upper threshold was applied to reject hot pixels, which are more likely due tosensor instabilities than to a real energy release. They are found to be not malfunctioning Mean 2.171Std Dev 0.7469 N u m be r o f p i x e l s Mean 2.171Std Dev 0.7469
Figure 5.
Distribution of the electronic noise of the sensor, estimated in images takenwith sensor in complete dark, and evaluated as the SD of the distribution of the countsfor each pixel. pixels since they disappear after a power cycle of the camera: therefore a dynamic (run-by-run) suppression is needed. They are efficiently identified as high-intensity, isolatedpixels, and distinguished by a true energy deposit, for which each pixel is surroundedby some other active pixels. A threshold is applied on the ratio R between the pixeland the average of the counts in a × pixels matrix surrounding it, and a minimumnumber of two pixels above noise in that matrix is required to discriminate good pixelsfrom hot ones. Only good pixels are retained for the subsequent clustering.The resolution of the resulting image is initially reduced by forming macro-pixels , byaveraging the counts in × pixel matrices. This is needed to reduce the combinatoricsof the subsequent clustering algorithm, in order to be executed in a reasonable timefor each image. On such × pixel map, a median filter [31] is applied, which iseffective in suppressing the electronics noise fluctuations in a × pixel matrix and it iscomputationally efficient, as described in more details in Ref. [32]. The output image ispassed to the basic clustering algorithm, described in the following. The basic clustering algorithm, called
Idbscan and described in details in Ref. [33],represents an improvement of the neighboring pixels clusters, called
Nnc , previouslyused to study the performance of the
Lemon detector with Fe radioactive source [30].It is briefly described also here, since it represents the seeding for the final clusteringalgorithm.The energy deposition in the sensitive volume of the TPC is estimated from the two-dimensional (2D) projection on the x – y axes of the light emitted in the multiplicationprocess within the GEMs planes. The pattern shows a large variation, depending on the0interacting particle. For images recorded with the Fe calibration source, the signatureof the typical 5.9 keV photons is a spot of few mm , with the exact size depending on thediffusion in the gas, i.e., on the distance from the anode, along z , of the point where theenergy release happens (see Fig. 4 left). Muons from cosmic rays travel across the gasvolume and leave a typical signature of a straight track, shown in Fig. 3 (right), but withseveral agglomeration with larger density along the path. Natural radioactivity showsan irregular pattern, sometimes curly, with several kinks along the path. Finally, thesignal from nuclear recoils due to neutrons, originated by the AmBe source, is expectedto be spot-like, or to emerge as short straight tracks with a length smaller than 1 mmfor energies below 100 keV, as shown by Fig. 2.Their track length and their size is found to depend a lot on the initial energy ofthe impinging neutron, and also on the mass of the recoiling nucleus in the He-CF gasmixture utilized in the Lemon detector.Thus, the clustering algorithm needs to be flexible enough to efficiently reconstructa diverse set of patterns, from small round spots to long and kinky tracks. A firststep of the clustering, called seeding , is used: it focuses in the clustering of spot-likeneighboring pixels. The method applied for the
Lemon detector is an evolution ofthe classic dbscan algorithm [34]. This is a non-parametric, density-based clustering,which groups together pixels above threshold with many neighbors, within a circle witha radius (cid:15) . Its distinctive characteristics making this method very suitable to the
Lemon case is its ability to label as outliers, and so not to include in the clusters, pixels that lieisolated in low-density regions, i.e., pixels from electronic noise of the sensor survivingthe zero suppression. The extension of dbscan used for
Lemon data analysis consistsin including a third dimension to the phase space of the points considered, adding tothe pixel position ( x – y coordinates) the measured number of photons in that pixel, N ph . This approach improves the combinatorial background rejection and the energyresolution with respect the previously used Nnc algorithm, as described in details inRef. [33].To be as inclusive as possible, and since different interactions may have vastlydifferent intensities, even varying along the track, the clustering procedure is iteratedthree times. First, the dbscan parameters were tuned to form clusters of dense (in x – y dimension) and intense (in the N ph dimension) pixels. The density in 3D is called sparsity . This step typically identifies either rare hot spots of the GEMs, or, efficiently,short nuclear recoils. The pixels belonging to the reconstructed clusters are then removedfrom the image, and the dbscan procedure is repeated, with looser sparsity parameters.The second iteration is tuned to efficiently reconstruct Fe round spots and slices oftracks from nuclear recoils with lower intensity. It also collects the agglomeration withlarger density along cosmic tracks, clearly visible in the example in Fig. 3 (right). Athird iteration of dbscan with even looser parameters is finally executed, targetingfaint portions of a cluster. These are especially used as a proxy for the characterizationof clustered noisy pixels, while the first two are used as seeds for the final clusteringstep, described in Sec. 4.3.1To be computationally viable, the
Idbscan basic clustering is performed on theimage with reduced resolution, 512 × Intel Xeon E5-2620
Python3 [35], andinterfaced with the CERN
Root6 v.6 [36].Examples of clustered pixels in two cases are shown in Fig. 6. The left panelshows an example of clusters reconstructed on the low-resolution image of one eventwith Fe source. Three spots are clearly visible: one, as typical for events with thiscalibration source with a moderate activity, is reconstructed by a single cluster of thesecond iteration. The other two are close enough that are merged in a single cluster of thesame iteration. The cases of merged spots containing twice the energy of a single X-raydeposit, given the activity of the Fe source, represent about one tenth of the clustersin this set of runs. The energy resolution is good enough to distinguish statisticallythe single and merged spots, as will be described in Sec. 4.4. The optimization of the
Idbscan parameters is done assuming a low pileup of events, typical of the runningconditions for a future underground run of the
Cygno project, of which
Lemon is theprototype, where the occurrences of such cases are expected to be negligible.The right panel shows the outcome of the
Idbscan algorithm on a longer track,presumably from natural radioactivity, and one possible short nuclear recoil. Thenuclear recoil candidate is very dense, highly-energetic, and isolated. Therefore, itis reconstructed as a single cluster in the first iteration. The long track shows severalclusters with higher intensity. One of them has a large energy, and it is reconstructedas an isolated single iteration-1 cluster. The rest of the track is reconstructed bymultiple iteration-2 clusters, which are split where the energy deposition has a minimumextending across too many pixels to be joined together in the same cluster. Events likethese, which are frequent for muons, natural radioactivity, but also signals from α -particles with higher energy, originating by the possible interaction of neutrons withthe plastic material of the field cage, justify the need of the subsequent step of the superclustering , which follows the track without splitting it in parts. This is describedin the following section. The aim of the superclustering procedure is to collect the majority of the pixels belongingto a track which is long and can be split in multiple parts in the clustering step describedbefore. Indeed, the main limitation of
Idbscan to follow a long track is mainlyoriginated by the non uniform energy release along the path length. As can be clearlyseen in Fig. 6 (right), or even in the example of a raw image of an event with two longcosmic rays in Fig. 3 (right), clusters with larger energy release are followed by regionsalong the path with a lower or even a zero release. These local minima are sometimesas large, in the 2D space, as the typical size of the (cid:15) parameter of dbscan [34]. Despitethe low electronic noise of the
ORCA-Flash 4.0 camera sensor, the energy releases in2
240 250 260 270 280 290 300 x (macro-pixels) y ( m a c r o - p i x e l s ) Rebinned image
240 260 280 300 320 340 360 380 400 x (macro-pixels) y ( m a c r o - p i x e l s ) Rebinned image
Figure 6.
Basic clusters reconstructed with the
Idbscan algorithm in the lowresolution (512 × Idbscan iteration. Left: clusters on spotsfrom Fe source, two of which are merged together. Right: Track from naturalradioactivity and a nuclear recoil candidate in an event with AmBe source. The longtrack is split in several basic clusters of different
Idbscan iteration. these local minima are similar in magnitude to the average single-pixel noise.The
Idbscan is limited in connecting the full length of an extended path, becauseof two reasons. First, inflating (cid:15) parameter as much as needed to cover the areas oflocal minima conflicts with the need to reject noise around the cluster. The basiccluster parameters were optimized for the
Lemon running conditions to collect mostof the signals with an energy as low as few keVs and to reject the typical noise of ≈ photon per pixel. This avoids collecting extra noise in the cluster, biasing the energyscale and worsening its resolution, and keeps the rate of fake clusters at a negligiblelevel [33]. Second, the iterative nature of the algorithm, with different parameters foreach iteration, each tuned for very different intensity, makes it convenient and efficient fora deposition of a fixed energy density (like the spots originating from the Fe source),but not for the cases as in Fig. 6 (right), where the same track is split in severalparts, some of them belonging to different iterations. This requires a method that cancontinuously follow the pattern of the track, profiting of the full resolution image, wherethe gradients of the energy deposition along the track trajectory are smaller than theones in the transverse direction, but still give information on the energy release pattern.Several existing algorithms were tested to profit of this, but executing any of them on thefull × image is not manageable CPU-wise, due to the huge pixel combinatorics.Therefore, the procedure adopted for the final supercluster reconstruction in the Lemon detector starts from defining the interesting regions in the image that may3contain pixels from an energy deposit. These are identified by the basic cluster algorithm
Idbscan previously described, which is applied on the × reduced-resolutionimage. In order to gather the peripheral pixels, especially along the track trajectorywhere breaks into small basic clusters may have happened, a window of × pixels isconsidered, around each pixel belonging to a macro-pixel clustered in a basic cluster.A full resolution image formed only by the interesting pixels passing the simple initialfiltering described in Sec. 4.1 is created. The gradients of the intensity N ph in suchimage are computed pixel-by-pixel to look for the edge region where the image turnsfrom signal to noise-only: ||∇ ( N ph ) || = (cid:115)(cid:18) ∂N ph ∂x (cid:19) + (cid:18) ∂N ph ∂y (cid:19) , (2)while the gradient direction is given by: θ = tan − (cid:18) ∂N ph ∂y / ∂N ph ∂x (cid:19) . (3)In order to reduce the effect of the noise, which induces fluctuations in the firstderivatives of Eq. 2, a Gaussian filter is applied, which smoothen the response byconvolving the pixel intensity with a Gaussian function, having as σ the SD of theintensities of all the pixels considered, and rejecting the ones falling outside a 5 σ window.The superclustering algorithm, applied on the filtered image, is an application of the morphological geodesic active contours [37,38], called Gac in the following. This methoduses an active contour finding, widely used in computer vision, where the boundary curve C of an object is detected by minimizing the energy E associated to C : E ( C ) = (cid:90) g ( N ph )( C ( p )) · |C p | dp, (4)where ds = |C p | dp is the arc-length parameterization of the curve in the 2D space, and g is the stopping edge function, which allows to select the boundary of the cluster. Inthe Gac method used for the
Lemon images, the g function is purely geometrical, anduses the geodesics of the image, i.e., the local minimal distance path joining points withthe same light intensity gradient. The function g ( N ph ) is given by: g ( N ph ) = 1 (cid:112) α |∇ G σ ∗ N ph | , (5)which is minimal in the edges of the image. The G σ ∗ N ph is the aforementioned σ Gaussian filter, and the parameter α , which regulates the strength of the filter, wastuned on typical Lemon images to be α = 100 .This method was chosen because it allows to follow patterns that may vary fromconvex to concave shape, eventually with kinks, e.g. in cases of δ -ray emissions. Toimprove the shrinking of the cluster boundary in the cases of tracks turning fromconcave to convex along their trail, the balloon force [38], which is a term added toEq. 4 to smooth the cluster contour, is set to -1, in order to push the contour towards aborder in the areas where the gradient is too small. A number of 300 iterations is usedto evolve the supercluster contour.4The example track shown in Fig. 6 (right) after the basic clustering step, is shownagain in full resolution, zoomed around the cluster, in Fig. 7 (left). The output ofthe superclustering with the Gac algorithm is shown on the right panel of the samefigure. The splitting of the cluster, happening at the basic clusters step, is recovered:the portions with high and low density along the path of the energy release are joinedtogether. Other three examples of superclustered images are shown in Fig. 8, in runswithout any artificial radioactive source. The top left panel shows an example of acosmic ray track fully reconstructed by the
Gac superclustering, which also includesa δ -ray in the middle of the track length. The top right panel shows an example ofcurly track from a candidate of natural radioactivity interaction; bottom panel showsan example where both a cosmic ray and a curly track are present. In this case, theextremes of the long and straight track are still split, but this is much rarer than afterthe basic clustering, and it happens when the local minima along the trajectories arecompatible with noise-only for more than ≈ x (pixels) y ( p i x e l s ) Image after zero suppression
240 260 280 300 320 340 360 380 400 x (macro-pixels) y ( m a c r o - p i x e l s ) Rebinned image
Figure 7.
Left: zoom on the full-resolution image of a track candidate in a run withthe AmBe radioactive source. Right: output of the superclustering on the rebinnedimage. The continuous line represents the approximate contour of the reconstructedsupercluster.
220 240 260 280 300 320 x (macro-pixels) y ( m a c r o - p i x e l s ) Rebinned image
300 310 320 330 340 350 360 370 x (macro-pixels) y ( m a c r o - p i x e l s ) Rebinned image
150 200 250 300 350 400 x (macro-pixels) y ( m a c r o - p i x e l s ) Rebinned image
Figure 8.
Superclusters reconstructed in a run without artificial radioactivesources. The continuous lines represent the approximate contours of the reconstructedsuperclusters. Top left: cosmic ray track fully reconstructed by the
Gac superclustering. A δ -ray is included in the supercluster. Top right: curly track from acandidate of natural radioactivity interaction. Bottom: a cosmic ray with the extremesnot joined to the main track, plus a curly track from natural radioactivity. Fe source
The containment of the energy in the supercluster was verified with simulations ofnuclear and electron recoils within the gas mixture of the
Lemon detector, performedwith
Srim [29]. For both types of recoils, for the energy range of interest for DM search,i.e., E (cid:46) keV, when considering deposits without electronics noise and no diffusionin the gas, the peak of the | E − E true | /E true distribution is within 5%. Adding a noiseapproximated as a Gaussian function with a mean and a SD equal to the ones observedin the pedestal runs, and a diffusion following the parameterization in Eq. 1, the fractionof the true energy contained in the supercluster decreases to about 80%. The decrease inthe energy containment in the supercluster is due to the smearing of the 2D track patternaround the periphery of the cluster, mostly due to the diffusion effect. This decreases thegradients in Eq. 2 around the edges, and so the supercluster can shrink more around thecrest, loosing part of the tails that can be confused more easily with the noise. A morerealistic noise description, and an improved diffusion model, based on the one measuredin data is necessary to tune the supercluster parameters in simulation to recover partof the containment. The energy resolution found in simulation (around 4%) is far fromthe measured one in data, around 18%, because of the absence, in the simulation, ofthe dominant contribution of the response fluctuations: Poissonian distribution of thenumber of primary electrons ionized in the gas and exponential behavior of the numberof secondary electrons produced in each GEM amplification stage [16]. Both of them areexpected to give rise to fluctuations of the order to 10%, that, once added in quadrature,can account for a large part of the measured energy resolution.The absolute energy scale was then calibrated with the energy distribution measuredin data with the Fe source, which provides monochromatic photons of 5.9 keV, withthe procedure described in Ref. [30]. The supercluster integral is defined as: I SC = cluster (cid:88) i N iph , (6)where N iph is the number of counts (photons) in the i th pixel, and the sum runs overall the pixels of the supercluster. While to perform the basic- and super-clustering onlypixels passing the zero suppression are considered, for the energy estimate in Eq. 6 allthe pixels within the cluster contours are counted, eventually having negative N iph afterthe pedestal subtraction. This choice is meant to avoid a bias on the energy estimate,since after the pedestal subtraction the distribution of the noise is centered around zero.The distribution of I SC , for a run taken in presence of Fe source, is shown in Fig. 9. Inaddition to the main peak, with a mean of about 2500 counts, a broader peak is clearlydistinguished, which represents the cases of two merged spots, with an integral twicethe single spot one. An example of such a merged spot is given in Fig. 6 (left). Theposition of the maximum in the single-spot distribution in runs with Fe source allowedto calibrate the absolute energy scale of the
Lemon detector. The energy resolutionfor the reconstructed
Gac superclusters is about 18%, similar to the one that can be7 (counts) SC I s upe r c l u s t e r s ( b k g s ub t r a c t ed ) – = 2541 m 3 counts – = 454 s
31 counts – = 5089 m 53 counts – = 786 s Figure 9.
Distribution of the supercluster integral, before the absolute energy scalecalibration is applied, in events with the Fe source. Clearly visible is the large peakof a single spot, and, at around twice the energy, a broader peak for the case of twoneighboring spots merged in a single supercluster. obtained with only the basic clustering step with
Idbscan [33], and improving the onewith the simple
Nnc algorithm previously used [30].Using runs with this monochromatic, high rate source, positioned at differentdistances from the GEM planes, a decrease of the light response for lower distancesfrom the GEM was observed. This effect is opposite to the expected behavior of alower light yield at larger distances. Indeed, it is expected that, during the drift alongthe z -direction, the ionization charge undergoes a diffusion in the TPC gas, and someelectrons are removed by attachment to the gas molecule. Consequently, some loss in thelight collection may be expected. The opposite behavior, instead, is clearly observed.While this effect is currently under study in more detail, it was attributed to a possiblesaturation effect of the GEMs, especially in the third stage of multiplication, wherethe charge density in one GEM hole is maximal. Under this hypothesis, an effective,empirical correction was developed, which relies on the charge density of a cluster froma Fe deposit. The light density, δ , is defined as: δ = I SC /n p , (7)where n p is the number of pixels passing the zero-suppression threshold (differentlyfrom the definition of I SC , where all the pixels in the supercluster are considered). Thiseffective calibration returns the absolute energy of a spot-like region, similar in sizeto the Fe clusters, as a function of the supercluster density, δ : E = c ( δ ) · I SC . Inthe hypothesis of saturation, the local density along the track is the parameter whichregulates the magnitude of the effect, thus the correction has to be applied dynamicallyfor slices of the supercluster having a size similar to the Fe spots. This is achieved8with the procedure described in the following.First, the supercluster skeleton , i.e., the 1-pixel-wide representation along the path,is reconstructed. This is achieved through a morphological thinning of the superclusterswith the iterative algorithm from Ref. [39,40]. Second, a pruning of the obtained skeletonis done, to remove residual small branches along the main pattern, using a hit-or-misstransform. The output of this process for one example track is shown in Fig. 10. For x (pixels) y ( p i x e l s ) Image after zero suppression x (pixels) y ( p i x e l s ) supercluster axis Figure 10.
Left: zoom on the full-resolution image of a track candidate in a run withthe AmBe radioactive source. Right: output of the skeletonization and pruning of thebranches for one example supercluster extended in space. the calibration procedure, the found skeleton is followed, starting from one of the twoend points, and circles having their center on a pixel of the skeleton and their radiusequal to the average spot size of the Fe clusters are defined. It was checked thatthis procedure includes all the pixels of the original cluster for the vast majority of theclusters considered. The local density δ s of the slice s is computed, and its integral I s is calibrated to an absolute energy through the effective correction E s = c ( δ s ) · I s .The pixels of the supercluster used for the slice calibration are removed (including theskeleton ones), and the procedure is iterated, until having included all the pixels. Thesum of the energies of all the slices is the estimate of the calibrated energy of thesupercluster: E SC = slices (cid:88) s E s (8)As a closure test of this procedure, the calibrated energy of the superclustersreconstructed in the runs with the Fe source is obtained. The value of the energypeak was obtained by fitting the distribution with the same function used in Fig. 9, andequals to m = 5 . ± . keV, compatible with the expected value. The calibration9procedure is an overkill for the case of the small Fe spots, but it is necessary forvery long cosmic ray tracks or even for medium-length superclusters from nuclear andelectron recoils. The energy resolution worsen after the calibration ( σ = 1 . ± . ,i.e., 25% energy resolution), as a sign that the empiric correction is still suboptimal.The skeletonization procedure provides a general method to estimate the tracklength ( l p ), accurate both in the case of straight and curving track. As a check, ithas been verified that, in the case of straight tracks, the length extracted in this waycoincides with the one of the major axis estimated with a singular value decomposition(SVD), described in the following section. For exactly round spots, the skeleton wouldcollapse in the center of the cluster and the resulting length would be 1 pixel, but thiscompletely symmetric case never happens in the considered samples.
5. Cluster shape observables
The interaction of different particles with the nuclei or the electrons in the gas of theTPC produces different patterns of the 2D projection of the initial 3D particle trajectory.These characteristics, to which we refer generically as cluster shapes observables , areuseful to discriminate different ionizing particles. In particular, they were used to selecta pure sample of nuclear recoil candidates produced by the interaction of the neutronsoriginating from the AmBe source and to identify various sources of backgrounds. Themain cluster shape observables are described in the following: • projected length and width: a SVD on the x × y matrix of the pixels belongingto the supercluster is performed. The eigenvectors found can be interpreted asthe directions of the two axes of an ellipse in 2D. The eigenvalues represent themagnitudes of its semiaxes: the major one is defined as length , l the minor one as width , w . These are well defined for elliptic clusters, or for long and straight tracks.The directions along the major and the minor axis are defined as longitudinal and transverse in the following. The longitudinal and transverse supercluster profiles,for the cosmic ray track candidates shown as an example in Fig. 8 (bottom) areshown in Fig. 11. The longitudinal profile shows the typical pattern of energydepositions in clusters, while the transverse profile, dominated by the diffusion inthe gas, shows a Gaussian shape. It has to be noted that the cluster sizes representonly the projection of the 3D track in the TPC on the 2D x – y plane; • projected path length: for curly and kinky tracks the values returned by the SVDof the supercluster are not an accurate estimates of their size. While the widthis dominated by the diffusion, the length for patterns like the one shown in theexample of Fig. 7 is ill-defined. Thus, the more general path length, l p , computedwith the skeletonization procedure in Fig. 10 is used to estimate the linear extentfor both straight and curved tracks. • Gaussian width: the original width of the track in the transverse direction isexpected to be much lower than the observed width induced by the diffusion in0 (mm) longitudinal X - N u m b e r o f pho t on s p e r s li ce (mm) transverse X N u m b e r o f pho t on s p e r s li ce Figure 11.
Supercluster profile in the longitudinal (top) or transverse (bottom)direction, for a long and straight cosmic ray track candidate shown in Fig. 8 (bottom).The longitudinal profile shows an energy deposition in sub-clusters, while the transversedirection shows the typical width of the diffusion in the gas. For the longitudinal profile,the line represent the average number of photons per slice. For the transverse profile,it represents a fit with a Gaussian PDF. the gas. Thus, as shown in Fig. 11 (right), the standard deviation, σ TGauss , canestimated by a fit with a Gaussian probability density function (PDF); • slimness: the ratio of the width over the path length, ξ = w/l p , represents theaspect ratio of the cluster. It is very useful to discriminate between cosmic rays-induced background (long and thin) from low energy nuclear or electron recoils(more elliptical or round, as the Fe spots); • integral: the total number of photons detected by all the pixels gathered in thesupercluster, I SC , as defined in Eq. 6; • pixels over threshold: the number of pixels in the supercluster passing the zero-suppression threshold, n p ; • density: the ratio δ of I SC , divided by n p , as defined in Eq. 7; • energy: the calibrated energy, expressed in keV. The calibration methodsimultaneously performs both the per-slice correction as a function of the local δ , and the absolute energy scale calibration, which corrects the non perfectcontainment of the cluster, i.e., the bias in the distribution of E/E true , using with Fe source.The projected supercluster path length, l p , and Gaussian transverse size, σ TGauss ,are shown in Fig. 12, for data taken in different types of runs. During the data-takingapproximately 3000 frames were recorded in absence of any external radioactive source( no-source sample). In these frames the interaction of ultra-relativistic cosmic rayparticles (mostly muons) are clearly visible as very long clusters. Internal radioactivityof the
Lemon materials also contribute with several smaller size clusters. About 1500frames were acquired with the AmBe source, and approximately calibration imageswith Fe source. In Fig. 12, as well as in the following ones showing other clusterproperties, the distributions obtained in runs without radioactive sources are normalizedto the AmBe data total CMOS exposure time. For the data with Fe source, since theactivity of the source is such to produce about 15 clusters/event, the data are scaled by1a factor one-tenth with respect to the AmBe exposure time for clearness. Considerationsabout the trigger efficiency scale factor between data with and without an radioactivesource are detailed later. The distributions in this section aim to show the differentcluster shape observables among the different kinds of events. (mm) p l c l u s t e r s ( no r m a li z ed t o A m B e ) AmBe no source · · · · (mm) p l b k g . s ub t r . da t a AmBe - no sourceFe - no source · · (mm) GaussT s c l u s t e r s ( no r m a li z ed t o A m B e ) AmBe no source · · · · (mm) GaussT s b k g . s ub t r . da t a AmBe - no sourceFe - no source · · Figure 12.
Supercluster sizes projected onto the x – y plane. Left: longitudinal pathlength, l p . Right: transverse Gaussian spread, σ TGauss . Filled points represent datawith AmBe source, dark gray (light blue) distribution represents data with Fe source(no source). The normalization of data without any radioactive source is scaled to thesame exposure time of the AmBe one. For the data with Fe source , a scaling factorof one tenth is applied for clearness, given the larger activity of this source.
Figure 12 shows the cluster sizes distributions in the longitudinal and transversedirections for different sets of runs. Data show an average Gaussian width for the Fe spots σ TGauss ≈ µ m (dominated by the diffusion in the gas), while it is larger,approximately 625 µ m, for data with AmBe source. The contribution of cosmic rays,present in all the data, is clearly visible in the data without any radioactive source,corresponding to clusters with a length similar to the detector transverse size (22 cm).Other observables are the slimness, ξ , and the light density, δ , shown in Fig. 13.The former is a useful handle to reject tracks from cosmic rays, which typically have aslim aspect ratio, i.e., low values of ξ , while the clusters from Fe are almost round,with values . (cid:46) ξ < . By construction, ξ < , since the width is computed alongthe minor axis of the cluster, and for round spots it peaks at around 0.9. The apparentthreshold effect is purely geometrical, due to the minimal size of the macro-pixel ( × )used at the basic clustering step which can be larger than a round spot from Fe. Datawith AmBe source, which contains a component of nuclear recoils, show a componentof round spots, similar in size to the ones of Fe, and a more elliptical component,with . < ξ < . values. Finally, the light density, δ , is the variable expectedto better discriminate among different candidates: cosmic rays induced background,electron recoils and nuclear recoil candidates. This is the variable used in this paper2for the final particle identification. The identification results can be improved usingadditional cluster shape variables, also profiting of their different correlations for signaland background clusters, via a multivariate approach, but here priority is given to thestraightforwardness, rather than the ultimate performance. x c l u s t e r s ( no r m a li z ed t o A m B e e v en t s ) AmBe no source · · · · x - - b k g . s ub t r . da t a AmBe - no sourceFe - no source · · (photons/pixel) d c l u s t e r s ( no r m a li z ed t o A m B e e v en t s ) AmBe no source · · · · (photons/pixel) d b k g . s ub t r . da t a AmBe - no sourceFe - no source · · Figure 13.
Supercluster variables. Left: slimness ξ ; right: light density δ . Filledpoints represent data with AmBe source, dark gray (light blue) distribution representsdata with Fe source (no source). The normalization of data without source is to thesame exposure time of the AmBe one. For the data with Fe, a scaling factor of onetenth is applied for clearness, given the larger activity of this source.
Finally, Fig. 14 (left) shows the calibrated energy ( E ) spectrum for thereconstructed superclusters. The energy spectrum shows the E = 5 . keV peak in thefirst bin of the distribution for data with Fe source, and the expected broad peakfor minimum ionizing particles traversing the ≈
20 cm gas volume at around 60 keV.The distribution of the observed average projected dEdl p for the no-source sample andfor the AmBe samples is shown in Fig. 14 (right). The broadening of the distributionis mainly due to the specific energy loss fluctuation in the gas mixture of the cosmicray particles. Its modal value, corrected for the effect of the angular distribution (anaverage inclination of 56 ◦ was measured from track reconstruction) is 2.5 keV/cm, ingood agreement with the Garfield prediction of 2.3 keV/cm.
The data with AmBe source, taken on the Earth surface, suffers from a large contributionof interactions of cosmic rays, and from ambient radioactivity, whose suppression is notoptimized for the
Lemon detector. The cluster shape observables provide a powerfulhandle to discriminate them from nuclear recoils candidates, but the small residualbackground needs to be statistically subtracted. The distributions shown earlier, wherethe different types of data are normalized to the same exposure time, demonstrate that3
E (keV) c l u s t e r s ( no r m a li z ed t o A m B e ) AmBe no source · · · · E (keV) b k g . s ub t r . da t a AmBe - no sourceFe - no source · · (keV/cm) p l dE/d c l u s t e r s ( no r m a li z ed t o A m B e ) AmBe no source · · · · (keV/cm) p l dE/d b k g . s ub t r . da t a AmBe - no sourceFe - no source · · Figure 14.
Supercluster calibrated energy spectrum (left) and their average dEdl p . Filledpoints represent data with AmBe source, dark gray (light blue) distribution representsdata with Fe source (no source). The normalization of data without source is to thesame exposure time of the AmBe one. For the data with Fe, a scaling factor of onetenth is applied for clearness, given the larger activity of this source. the live-time normalization provides already a good estimate of the amount of cosmicrays in data with radioactive sources. This approach does not account for a possible biasfrom the trigger, which is generated by the PMT signals, as described in Sec. 2. Indeed,in runs with the AmBe source, the PMT can trigger both on signals from neutron recoilsor photons produced by the
Am, and on ubiquitous signals from cosmic rays, whilein the sample without source only the latter are possible. Therefore, during the sameexposure time, the probability to trigger on cosmic rays is lower in events with AmBethan in no-source events. The trigger efficiency scale factor, ε SF , can be obtained asthe ratio of the number of clusters selected in pure control samples of cosmic rays ( CR )obtained on both types of runs: ε SF = N AmBeCR N no − sourceCR . (9)The CR control region is defined by selecting clusters with l > cm, ξ < . , σ TGauss < mm, and having an energy within a range dominated by the cosmic rayscontribution, < E < keV. The selected clusters show small values of δ ≈ , wellcompatible with the small specific ionization of ultra-relativistic particles. This sampleis limited in statistics, but it is expected to be almost 100% pure. The scale factorobtained is ε SF = 0 . ± . .In Fig. 15 the typical light density and polar angle (with respect the horizontalaxis) distributions for long clusters of any energy, still dominated by cosmic rays, areshown for the AmBe and for the no-source sample, after having applied the ε SF scalefactor to the latter. Clusters with δ < are thus expected to be mostly coming frommuon tracks, and they show indeed a polar angle which is shifted at values towards ◦ .4 (photons/pixel) d c l u s t e r s ( no r m a li z ed t o A m B e e v en t s ) (photons/pixel) d - - b k g . s ub t r . da t a AmBe - no sourceAmBe - no source (deg) q c l u s t e r s ( no r m a li z ed t o A m B e e v en t s ) (deg) q - b k g . s ub t r . da t a AmBe - no sourceAmBe - no source
Figure 15.
Supercluster light density δ (left) and polar angle (right) - with respectthe horizontal axis - distributions for long clusters, dominated by cosmic rays tracks.Filled points represent data with AmBe source, light blue distribution represents datawithout any radioactive source. The normalization of data without source is to thesame exposure time of the AmBe one, accounting for the trigger scale factor ε SF , asdefined in the text.
6. Nuclear recoil identification results
As mentioned in the previous Section, the 1D observable chosen to distinguish the signalof nuclear recoils from the various types of background is the energy density δ of thecluster. To enhance the purity of the signal sample, a preselection was applied, prior to a tighterselection on δ : clusters with l p > . cm or ξ < . were rejected to primarily suppressthe contribution from cosmic rays. A further loose requirement δ > photons/pixel wasalso applied to remove the residual cosmic rays background based on their low specificionization. These thresholds, which only reject very long and narrow clusters, are veryloose for nuclear recoils with E < MeV energies, given the expected range in simulatedevents, shown in Fig. 2, of less than 1 cm. Thus the preselection efficiency for signalis assumed to be 100%. For electron recoils it can be estimated on data by using the Fe data sample, and is measured to be ε preselB = 70% . Since the X-ray photo-electronsof this source are monochromatic, the estimate of the electron recoils rejection is onlychecked for an energy around E = 5 . keV. The spectrum of nuclear recoils from AmBesource, instead, extends over a wider range of energies, around [1–100] keV.With this preselection, the distribution in the 2D plane δ – l p is shown in Fig. 16 forAmBe source and no-source data and for the resulting background-subtracted AmBedata. The latter distribution shows a clear component of clusters with short length5( l p (cid:46) cm) and high density ( δ (cid:38) ), expected from nuclear recoils deposits.In addition, it shows a smaller component, also present only in the data with AmBesource, of clusters with a moderate track length, . (cid:46) l p (cid:46) . cm, and a lower energydensity than the one characteristic of the nuclear recoils ( (cid:46) δ (cid:46) ). Since the densityis inversely proportional to the number of active pixels n p , which is correlated to thetrack length, the almost linear decrease of δ as a function of l p points to a componentwith fixed energy. The Am is expected to produce photons with E = 59 keV. Thishypothesis is verified by introducing an oblique selection in the δ − l p plane: | δ − y | < ,where y = 14 − p l / , for the clusters with < l p < pixels, defining the photoncontrol region , P R . The approximate oblique region in the δ − l p plane corresponding to P R is also shown in Fig. 16. The obtained energy spectrum for these clusters is shownin Fig. 17, which indeed shows a maximum at E = 60 . ± . keV, within the expectedresolution. These events are thus rejected from the nuclear recoils candidates by vetoingthe P R phase space. p l51015202530 ( pho t on s / p i x ) δ (AmBe) p l51015202530 ( pho t on s / p i x ) δ (no source) p l51015202530 ( pho t on s / p i x ) δ (no source) × (AmBe) - 0.75 Figure 16.
Supercluster light density δ versus length l p , for data with AmBe source(left), data without any artificial source (middle), and the resulting background-subtracted AmBe data. The normalization of data without source is to the sameexposure time of the AmBe one, accounting for the trigger scale factor ε SF , as definedin the text. The orange ellipse represents the approximate contour of the 59 keVphotons control region ( P R ) defined in the text.
An independent information to the light detected by the sCMOS sensor of the camerais obtained from the PMT pulse, used to trigger the image shooting. For each imageacquired, the corresponding PMT pulse waveform is recorded. Tracks from cosmic rays,which typically have a large angle with respect the cathode plane, as shown in Fig. 15(right), show a broad PMT waveform, characterized by different arrival times of theseveral ionization clusters produced along the track at different z . Conversely, spot-likesignals like Fe deposits or nuclear recoils are characterized by a short pulse, as shownin Fig. 18.6
E (keV) c l u s t e r s ( no r m a li z ed t o A m B e ) AmBe no source · · E (keV) b k g . s ub t r . da t a – mean = 60.9 1.9 keV – = 15.3 s AmBe - no sourceAmBe - no source
Figure 17.
Calibrated energy spectrum for candidates in the control region
P R ,defined in the text. The background-subtracted distribution is fitted with a GaussianPDF, which shows a mean value compatible with E = 59 keV originated from the Am γ s interaction within the gas. Figure 18.
Example of two acquired waveforms: one short pulse recorded in presenceof Fe radioactive source, together with a long signal very likely due to a cosmic raytrack.
The Time Over Threshold (
TOT ) of the PMT pulse was measured, and is shown inFig. 19. It can be seen from the region around 270 ns, dominated by the cosmic rays also7in the data with the AmBe source, that the trigger scale factor ε SF also holds for thePMT event rate. As expected, spot-like clusters (in 3D) correspond to a short pulse in PMT T.O.T. (ns) c l u s t e r s ( no r m a li z ed t o A m B e ) AmBe no source · · · · PMT T.O.T. (ns) - - - - b k g . s ub t r . da t a AmBe - no sourceFe - no source · · Figure 19.
PMT waveform time over threshold (
T OT ). The last bin integratesall the events with
T OT > ns. Filled points represent data with AmBe source,dark gray (light blue) distribution represents data with Fe source (no source). Thenormalization of data without source is to the same exposure time of the AmBe one,with trigger scale factor ε SF applied. For the data with Fe, a scaling factor of onetenth is applied for clearness, given the larger activity of this source. the PMT, while cosmic ray tracks have a much larger pulse. The contribution of cosmicray tracks is clearly visible in the data with radioactive sources. A selection on thisvariable is helpful to further reject residual cosmic rays background present in the AmBeor Fe data, in particular tracks which may have been split in multiple superclusters,like the case shown in Fig. 8 (bottom), and thus passing the above preselection on thecluster shapes. A selection
T OT < ns is then imposed. It has an efficiency of 98% oncluster candidates in AmBe data (after muon-induced background subtraction), whileit is only 80% efficient on data with Fe source. This larger value is expected becauseof the residual contamination of signals from cosmic rays, which fulfill the selectionbecause their track is split in multiple sub-clusters, or because they are only partiallyvisible in the sCMOS sensor image. These can be eventually detected as long, in the timedimension, by the PMT. The light density and the energy spectrum of the preselectedclusters are shown in Fig. 20. Fe events rejection
The light density distribution, after the above preselection and cosmic ray suppression,appear to be different among the data with AmBe source, data with Fe source, anddata without any artificial source. The cosmic-background-subtracted distributionsof δ in AmBe data and Fe data, shown in the bottom panel of Fig. 20 (left), are8 (photons/pixel) d c l u s t e r s ( no r m a li z ed t o A m B e ) AmBe no source · · · · (photons/pixel) d b k g . s ub t r . da t a AmBe - no sourceFe - no source · · E (keV) - - -
10 110 c l u s t e r s / b i n s i z e ( no r m a li z ed t o A m B e ) AmBe no source · · · · E (keV) - - -
10 110 b k g . s ub t r . da t a AmBe - no sourceFe - no source · · Figure 20.
Supercluster light density δ (left) and calibrated energy E (right), afterthe preselection and cosmic ray suppression described in the text to select nuclearrecoil candidates. Filled points represent data with AmBe source, dark gray (lightblue) distribution represents data with Fe source (no-source). The normalization ofno-source data is to the same exposure time of the AmBe data, with the trigger scalefactor ε SF applied. For the data with Fe, a scaling factor of one tenth is applied forclearness, given the larger activity of this source. used to evaluate a curve of 5.9 keV electron recoils rejection ( − ε δB ) as a functionof signal efficiency ( ε δS ), obtained varying the selection on δ , shown in Fig. 21. Thesame procedure could be applied to estimate the rejection factor against the cosmic rayinduced background, but this is not shown because of the limited size of the no-sourcedata. This kind of background will however be negligible when operating the detectorunderground, in the context of the Cygno project, so no further estimates are givenfor this source.Table 1 shows the full signal efficiency and electrons rejection factor for two exampleworking points, WP and WP , having 40% and 50% signal efficiency for the selectionon δ , averaged over the full energy spectrum exploited in the AmBe data. Theycorrespond to a selection δ > and δ > , respectively. While this cut-based approachis minimalist, and could be improved by profiting of the correlations among δ and theobservables used in the preselection in a more sophisticated multivariate analysis, itshows that a rejection factor approximately in the range [ − - − ] of electron recoilsat E = 5 . keV with a gaseous detector at atmospheric pressure can be obtained, whileretaining a high fraction of signal events. The energy spectrum for the candidates with ε totalS =50% in the AmBe sample is shownin Fig. 22 (left). The signal efficiency is then computed for both the example working9 ) d S e Signal efficiency ( ) d B e B a ck g r ound r e j e c t i on ( - > 10 photons/pixel d > 11 photons/pixel d Figure 21.
Background rejection as a function of the signal efficiency, varying theselection on the δ variable in data with either Fe (background sample) or AmBe(signal sample) sources.
Table 1.
Signal (nuclear recoils induced by AmBe radioactive source) and background(photo-electron recoils of X-rays with E = 5 . keV from Fe radioactive source)efficiency for two different selections on δ . working point Signal efficiency Background efficiency ε preselS ε δS ε totalS ε preselB ε δB ε totalB WP WP ε totalB , represents a γ backgroundefficiency at a fixed energy E = 5 . keV, i.e., the energy of the photons emitted by the Fe source. For the WP , the efficiency for very low-energy recoils, E = 5 . keV, isstill 18%, dropping to almost zero at E (cid:46) keV.Two candidate nuclear recoils images, fulfilling the WP selection (with a lightdensity δ (cid:38) photons/pixels and with energies of 5.2 and 6.0 keV) are shown inFig. 23. The displayed images are a portion of the full-resolution frame, after thepedestal subtraction. While the determination of the direction of detected nuclear recoilis still under study, it appears pretty clear from the image that some sensitivity to theirdirection, even at such low energies, is retained and can be further exploited.0 E (keV) - - -
10 110 c l u s t e r s / b i n s i z e ( no r m a li z ed t o A m B e ) AmBe no source · · · · E (keV) - - -
10 110 b k g . s ub t r . da t a AmBe - no sourceFe - no source · · E (keV) ) t o t a l S e s i gna l e ff i c i en cy ( =4% totalB e =1% totalB e Figure 22.
Left: supercluster calibrated energy E (left), after the full selection, whichincludes δ > , 50% efficient on signal, to select nuclear recoil candidates. Filled pointsrepresent data with AmBe source, dark gray (light blue) distribution represents Fesource (no-source) data. The normalization of no-source data is to the same exposuretime of the AmBe data, with the trigger scale factor ε SF applied. For the Fe data,a scaling factor of one tenth is applied for clearness, given the larger activity of thissource. Right: efficiency for nuclear recoil candidates as a function of energy, estimatedon AmBe data, for two example selections, described in the text, having either 4% or1% efficiency on electron recoils at E = 5 . keV.
660 680 700 720 740 760 x (pixels) y ( p i x e l s ) Image after zero suppression x (pixels) y ( p i x e l s ) Image after zero suppression
Figure 23.
Examples of two nuclear recoil candidates, selected with the full selection,shown in a portion of × pixel matrix, after the zero suppression of the image.Left: a candidate with E = 5 . keV and δ = 10 . , right: a candidate with E = 6 . keVand δ = 10 .
7. Conclusion and outlook
A method to efficiently identify recoiling nuclei after an elastic scattering with fastneutrons with an optically readout TPC was presented in this paper. A 7 liter prototypewas employed by exposing its sensitive volume to two kinds of neutral particles in anoverground location: • photons with energy of 5.9 keV and 59 keV respectively provided by a radioactivesource of Fe and by one of
Am able to produce electron recoils with equalenergy by means of photoelectric effect; • neutrons with kinetic energy of few MeV produced by an AmBe source that cancreate nuclear recoils with kinetic energy lower than the neutron ones.The high sensitivity of the adopted sCMOS optical sensor allowed a very goodefficiency in detecting events with an energy released in gas even below 10 keV.Moreover, the possibility of exploiting the topological information (shape, size andmore) of clusters of emitted light allowed to develop algorithms able to reconstruct notonly the total deposited energy, but also to identify the kind of the recoiling ionizingparticles in the gas (either an electron or a nucleus). Cosmic ray long tracks are alsoclearly separated.Because of their larger mass and electric charge, nuclear recoils are expected torelease their energy by ionizing the gas molecules in few hundreds µ m while the electronsare able to travel longer paths. For this reason, by exploiting the spatial distribution ofthe collected light, it was possible to identify 5.9 keV electron recoils with an efficiencyof 96.5% (99.2%) against nuclear recoils by retaining a capability of detecting them withan efficiency of 50% (40%), averaged across the measured AmBe spectrum.In particular, the nuclear recoil detection efficiency was measured to be 40% fordeposited energies lower than 20 keV and 14% in the range (5–10) keV.The results obtained in the studies presented in this paper can be improvedby means of more sophisticated analyses exploiting a multivariate approach, whichcombines a more complete topological information about the light distribution alongthe tracks. Additional enhancement of sensitivity can be achieved with a DAQ systemcollecting single PMT waveforms to be correlated with the track reconstructed in thesCMOS images.
8. Acknoledgements
We are grateful to Servizio Sorgente LNF... This work was supported by the EuropeanResearch Council (ERC) under the European UnionâĂŹs Horizon 2020 research andinnovation program (grant agreement No 818744)âĂİ. [1] B. W. Lee and S. Weinberg, “Cosmological lower bound on heavy-neutrino masses,”
Phys. Rev.Lett. , vol. 39, pp. 165–168, Jul 1977.[2] T. M. Undagoitia and L. Rauch, “Dark matter direct-detection experiments,”
Journal of PhysicsG: Nuclear and Particle Physics , vol. 43, p. 013001, dec 2015. [3] Y. Hochberg, E. Kuflik, T. Volansky, and J. G. Wacker, “Mechanism for thermal relic dark matterof strongly interacting massive particles,” Phys. Rev. Lett. , vol. 113, p. 171301, Oct 2014.[4] D. E. Kaplan, M. A. Luty, and K. M. Zurek, “Asymmetric dark matter,”
Phys. Rev. D , vol. 79,p. 115016, Jun 2009.[5] K. PETRAKI and R. R. VOLKAS, “Review of asymmetric dark matter,”
International Journal ofModern Physics A , vol. 28, no. 19, p. 1330028, 2013.[6] K. M. Zurek, “Asymmetric dark matter: Theories, signatures, and constraints,”
Physics Reports ,vol. 537, no. 3, pp. 91 – 121, 2014. Asymmetric Dark Matter: Theories, signatures, andconstraints.[7] F. Mayet, A. Green, J. Battat, J. Billard, N. Bozorgnia, G. Gelmini, P. Gondolo, B. Kavanagh,S. Lee, D. Loomba, J. Monroe, B. Morgan, C. OâĂŹHare, A. Peter, N. Phan, and S. Vahsen,“A review of the discovery reach of directional dark matter detection,”
Physics Reports , vol. 627,pp. 1 – 49, 2016. A review of the discovery reach of directional Dark Matter detection.[8] E. Baracchini, G. Cavoto, G. Mazzitelli, F. Murtas, F. Renga, and S. Tomassini, “Negativeion time projection chamber operation with SF at nearly atmospheric pressure,” Journal ofInstrumentation , vol. 13, pp. P04022–P04022, Apr. 2018.[9] M. Marafini, V. Patera, D. Pinci, A. Sarti, A. Sciubba, and E. Spiriti, “ORANGE: A high sensitivityparticle tracker based on optically read out GEM,”
Nucl. Instrum. Meth. , vol. A845, pp. 285–288,2017.[10] V. C. Antochi, E. Baracchini, G. Cavoto, E. D. Marco, M. Marafini, G. Mazzitelli, D. Pinci,F. Renga, S. Tomassini, and C. Voena, “Combined readout of a triple-GEM detector,”
JINST ,vol. 13, no. 05, p. P05001, 2018.[11] D. Pinci, E. Di Marco, F. Renga, C. Voena, E. Baracchini, G. Mazzitelli, A. Tomassini, G. Cavoto,V. C. Antochi, and M. Marafini, “Cygnus: development of a high resolution TPC for rare events,”
PoS , vol. EPS-HEP2017, p. 077, 2017.[12] G. Mazzitelli, V. A. Antochi, E. Baracchini, G. Cavoto, A. De Stena, E. Di Marco, M. Marafini,D. Pinci, F. Renga, S. Tomassini, and C. Voena, “A high resolution TPC based on GEMoptical readout,” in , pp. 1–4, Oct 2017.[13] D. Pinci, E. Baracchini, G. Cavoto, E. Di Marco, M. Marafini, G. Mazzitelli, F. Renga,S. Tomassini, and C. Voena, “High resolution TPC based on optically readout GEM,”
NuclearInstruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectorsand Associated Equipment , 2018.[14] I. A. Costa, E. Baracchini, R. Bedogni, F. Bellini, L. Benussi, S. Bianco, M. Caponero, G. Cavoto,E. D. Marco, G. D’Imperio, G. Maccarrone, M. Marafini, G. Mazzitelli, A. Messina, F. Petrucci,D. Piccolo, D. Pinci, F. Renga, G. Saviano, and S. Tomassini, “CYGNO: Triple-GEM opticalreadout for directional dark matter search,”
Journal of Physics: Conference Series , vol. 1498,p. 012016, apr 2020.[15] C. Collaboration, “CYGNO Phase-I: One cubic meter demonstrator.” https://web.infn.it/cygnus/cygno , 2020.[16] D. Pinci,
A triple-GEM detector for the muon system of the LHCb experiment . PhD thesis, CagliariUniversity, CERN-THESIS-2006-070, 2006.[17] Hamamatsu,
ORCA-Flash4.0 V3 Digital CMOS camera , 2018. .[18] M. Marafini, V. Patera, D. Pinci, A. Sarti, A. Sciubba, and E. Spiriti, “High granularity trackerbased on a Triple-GEM optically read by a CMOS-based camera,”
JINST , vol. 10, no. 12,p. P12010, 2015.[19] HZC Photonics,
XP3392 Photomultiplier . .[20] V. C. Antochi, G. Cavoto, I. Abritta Corsta, E. Di Marco, G. D’Imperio, F. Iacoangeli, M. Marafini,A. Messina, D. Pinci, F. Renga, C. Voena, E. Baracchini, A. Cortez, G. Dho, L. Benussi, S. Bianco, C. Capoccia, M. Caponero, G. Maccarrone, G. Mazzitelli, A. Orlandi, E. Paoletti,L. Passamonti, D. Piccolo, D. Pierluigi, F. Rosatelli, A. Russo, G. Saviano, S. Tomassini, R. A.Nobrega, and F. Petrucci, “A GEM-based optically readout time projection chamber for chargedparticle tracking,” 2020.[21] R. Campagnola, “Study and optimization of the light-yield of a triple-GEM detector ,” Master’sthesis, Sapienza University of Rome, 2018.[22] N. Torchia, “Development of a tracker based on GEM optically readout ,” Master’s thesis, SapienzaUniversity of Rome, 2016.[23] CAEN, , 2017. .[24] R. Veenhof, “GARFIELD, recent developments,”
Nucl. Instrum. Meth. A , vol. 419, pp. 726–730,1998.[25] R. Veenhof, “GARFIELD, a drift chamber simulation program,”
Conf. Proc. C , vol. 9306149,pp. 66–71, 1993.[26] G. Mazzitelli, A. Ghigo, F. Sannibale, P. Valente, and G. Vignola, “Commissioning of theDA φ NE beam test facility,”
Nuclear Instruments and Methods in Physics Research Section A:Accelerators, Spectrometers, Detectors and Associated Equipment , vol. 515, no. 3, pp. 524 – 542,2003.[27] I. Abritta Costa, E. Baracchini, F. Bellini, L. Benussi, S. Bianco, M. A. Caponero, G. Cavoto,G. D’Imperio, E. Di Marco, G. Maccarrone, M. Marafini, G. Mazzitelli, A. Messina, F. Petrucci,D. Piccolo, D. Pinci, F. Renga, F. Rosatelli, G. Saviano, and S. Tomassini, “Stability anddetection performance of a GEM-based optical readout TPC with He/CF gas mixtures,” Journal of Instrumentation , vol. xx, p. xxx, jul 2020.[28] S. Agostinelli et al. , “GEANT4—a simulation toolkit,”
Nucl. Instrum. Meth. A , vol. 506, p. 250,2003.[29] J. Ziegler, “Srim – 2003,”
Nuclear Instruments and Methods in Physics Research Section B: BeamInteractions with Materials and Atoms , vol. 219-220, no. 3, pp. 1027–1036, 2004.[30] I. Abritta Costa, E. Baracchini, F. Bellini, L. Benussi, S. Bianco, M. A. Caponero, G. Cavoto,G. D’Imperio, E. Di Marco, G. Maccarrone, M. Marafini, G. Mazzitelli, A. Messina, F. Petrucci,D. Piccolo, D. Pinci, F. Renga, F. Rosatelli, G. Saviano, and S. Tomassini, “Performance ofoptically readout GEM-based TPC with a Fe source,”
Journal of Instrumentation , vol. 14,pp. P07011–P07011, jul 2019.[31] Y. Dong and S. Xu, “A new directional weighted median filter for removal of random-valued impulsenoise,”
IEEE Signal Processing Letters , vol. 14, no. 3, pp. 193–196, 2007.[32] G. S. P. Lopes, E. Baracchini, F. Bellini, L. Benussi, S. Bianco, G. Cavoto, I. A. Costa, E. Di Marco,G. Maccarrone, M. Marafini, G. Mazzitelli, A. Messina, R. A. Nobrega, D. Piccolo, D. Pinci,F. Renga, F. Rosatelli, D. M. Souza, and S. Tomassini, “Study of the impact of pre-processingapplied to images acquired by the cygno experiment,” in
Pattern Recognition and Image Analysis (A. Morales, J. Fierrez, J. S. Sánchez, and B. Ribeiro, eds.), (Cham), pp. 520–530, SpringerInternational Publishing, 2019.[33] I. Abritta et al. , “A density-based clustering algorithm for the cygno data analysis,”
In preparation ,vol. 00, no. 0, pp. 00–00, 2020.[34] M. Ester, H.-P. Kriegel, J. Sander, and X. Xu, “A density-based algorithm for discovering clustersin large spatial databases with noise,” pp. 226–231, AAAI Press, 1996.[35] G. Van Rossum and F. L. Drake,
Python 3 Reference Manual . Scotts Valley, CA: CreateSpace,2009.[36] R. Brun and F. Rademakers, “ROOT: An object oriented data analysis framework,”
Nucl. Instrum.Meth. A , vol. 389, pp. 81–86, 1997.[37] V. Caselles, R. Kimmel, and G. Sapiro, “Geodesic Active Contours,”
International Journal ofComputer Vision , vol. 22, pp. 61–79, 1997.[38] P. MÃąrquez-Neila, L. Baumela, and L. Alvarez, “A morphological approach to curvature- based evolution of curves and surfaces,” IEEE Transactions on Pattern Analysis and MachineIntelligence , vol. 36, no. 1, pp. 2–17, 2014.[39] Z. Guo and R. W. Hall, “Parallel thinning with two-subiteration algorithms,”
Commun. ACM ,vol. 32, p. 359âĂŞ373, Mar 1989.[40] L. Lam, S. Lee, and C. Y. Suen, “Thinning methodologies - a comprehensive survey,”