OpHit Slicing Dan Pershey Feb 11, 2019
Overview ❑ Implemented an OpHit clusterer, based on DBScan, but slightly modified ❑ Radiological OpHit’s are infrequent enough in our PDS that algorithm measures delayed scintillation light in LAr • Exciting moving forward looking at this in the context of calorimetry, PID ❑ Current algorithm, simplest procedure you could write down that works: • Sort OpHit’s by time • Scan in time and calculate total PE in each time window • If total PE > some threshold, read out the whole PDS for the time window (up to 0.5 μ s) 2
Helpful Numbers – how much radiological bkg do we expect ❑ At a 43 cm 2 ARAPUCA SiPM design, our simulation predicts about 4000 distinct interactions from the radio sim that make some blip in the PDS per drift window • At a drift window of 4.4 ms , that’s right around 1 radiological / μ s • Almost all decays are from 39 Ar, recording 1- 2 PE’s / decay ❑ A rough cut of 400(300) cm around a reconstructed flash center is visible to about 20(12)% of the detector • 0.7 x π x (400) 2 / 1400 / 1200 – where 0.7 is an edge effect fudge factor ❑ So, a hard cut of 400 cm around a reco’d vertex will slice in light from 0.2 radiological decays / μ s of scanning time ❑ Prompt light from a typical SNB neutrino’s OpFlash is spread over 0.25 μ s – we expect 0.25 μ s x 2PE/decay x 0.2decay/ μ s = 0.1 PE contamination per event ❑ At 0.5 PE/MeV, that’s a 2% contribution to a 10 MeV neutrino ❑ Low enough, we can hope to include delayed scintillation light 3
Example Op Flash Evolution E ν = 12.75 MeV 0 < t < 1 μ s Neutrino Light Radiologicals Y (cm) Y (cm) Y (cm) Y (cm) Z (cm) Z (cm) 4
Example Op Flash Evolution E ν = 12.75 MeV 1 < t < 2 μ s Neutrino Light Radiologicals Y (cm) Y (cm) Z (cm) Z (cm) 5
Example Op Flash Evolution E ν = 12.75 MeV 2 < t < 3 μ s Neutrino Light Radiologicals Y (cm) Y (cm) Z (cm) Z (cm) 6
Example Op Flash Evolution E ν = 12.75 MeV 3 < t < 4 μ s Neutrino Light Radiologicals Y (cm) Y (cm) Z (cm) Z (cm) 7
Example Op Flash Evolution E ν = 12.75 MeV 4 < t < 5 μ s Neutrino Light Radiologicals Y (cm) Y (cm) Z (cm) Z (cm) 8
Classifying Sources of Background Hits ❑ Almost all unwanted OpHits have either 1 or 2 PE’s ❑ 39 Ar is the largest single contributor ❑ There’s a large portion of the hits that are not associated with any physics • These are from the noise simulation • Noise hits are often correlated with physics • Cross-talk is simulated, so that neutrinos induce statistical noise in our PD’s ❑ So, I take it that cross-talk is in-separable from our neutrino OpHits, but they count against OpHit cluster purity 9
Clustering Algorithm – DBSCan Look up DBScan on ❑ Relies on a metric to calculate the distance Wikipedia between each pair of events in your space • For us, this is a function of Δ t, Δ z, and Δ y ❑ Cycle through each point in your space, • Find how many other pts are in the neighborhood within 1 unit of test pt • If there are at least N 0 pts in neighborhood, this pt belongs to a slice, and is labeled a “core” pt • Add all neighborhood pts to the cluster • If a point is in the neighborhood of a core pt, but does not have N 0 pts in its neighborhood, it is a peripheral event • Added to the cluster, but any extra pts in its neighborhood are not ❑ N 0 technically a free parameter, but fix it to 3 • Similar to putting a 3PE threshold on OpFlashes, going lower would increase our data rate 10
Modifying DBScan ❑ DBScan works great in the context it’s designed for, but doesn’t do great at slicing in the context of a space with multiple “slice densities” ❑ Fails miserably with accurately reconstructed both prompt and delayed hits into one cluster • Could not massage parameters to accept all prompt+delayed light in same clusters at multiple energies without including a bunch of radiological decays ❑ Perform DBScan to cluster the prompt light – with restrictions • While iterating through core DBScan hits, also calculate a cluster centroid – the hit with the highest density of points surrounding it • Do not include extra core hits if its distance is more than 2 units away from the core centroid ❑ Then, slice in the delayed light by accepting hits within 5 μ s and R scale cm away from the centroid 11
Slicer Performance ❑ We have a well-defined FOM for how well a slicing algorithm does • For each event, calculate purity x completeness • Purity = fraction of clustered light that came from neutrino • Completeness = fraction of light deposited by the neutrino captured by the slice • FOM = the mean of the distribution of this product for a set of MC ❑ Tested performance on a range of algorithm parameters (next slide) • Chose length scale = 400 cm and time scale = 0.15 μ s 12
Parameter Tuning ❑ Two parameters to tune – the length scales for distance and time that go into the metric ❑ No clear winner, distributions are relatively flat, independent of E ν • Seems to be a small peak in R scale curves, set scale = 400 cm • Set T scale = 0.15 μ s, roughly half of the prompt time constant 13
❑ Most events have prompt(delayed) purity > 90(80)% ❑ Between the two time windows, we can usually capture all the light released E ν = 15.25 MeV 14
Prompt and Delayed Purity ❑ Both shapes are similar, and tied to PE ❑ Delayed purity is lower – seems to be from increased cross-talk OpHit’s that light up slightly delayed from the prompt flash E ν = 15.25 MeV 15
Completeness Also Tied to Total PE Observed ❑ Generally speaking, we capture almost all of the light deposited for neutrinos that deposited > 200 PE in the detector • At a mean 20 PE/cm, that’s a 10 MeV neutrino ❑ Though, second population at total PE deposited > 400 and completeness < 0.8 ❑ Don’t currently understand what these are, but would be interesting E ν = 15.25 MeV 16
Delayed Scintillation Light + Calorimetry ❑ Compare the fraction of neutrino light reco’d from the prompt window and prompt+delayed windows • Prompt window: find the time with highest density of PE in the flash, and go +- 250 ns from that time • Similar to current OpFlash reco, which looks for an up to 500 ns window with activity ❑ Clearly there is more variance if you only consider prompt light 17
Summary ❑ Implemented a slicing algorithm for OpHits ❑ Algorithm tailored to best fit out fast and slow time components – which will have different density of hits at different neutrino energies • Potentially very interesting to propagate this to simulation and analysis • There’s a lot of information in the prompt + delayed total and ratio • Can help with calorimetry, PID with rejecting radiologicals 18
Recommend
More recommend
